This adds new simulation configurations that more closely match the SLACS HST images. Sebastian started this work, the first commit of this PR is just me copying his files into the repo.
The new configs are in the acs folder. Besides the main config_train.py and config_val.py, I added three pairs of configs that gradually relax the assumptions back to those of the paltas paper:
noll represents no lens light
noll_fixz also fixes the source and lens redshifts
fantasy improves the exposure time and resolution
Changes with respect to the paltas paper configuration:
Time per exposure changed from 1320 seconds to Norm(mean=2184/4, std=116/4) seconds. Note we still assume four individual exposures, so this distribution is sampled once, and the total integrated exposure time is four times what we draw from this. This then matches the distribution (2184 +- 116 sec) of SLACS exposure times for the full-orbit observations (at least in mean and std; the min and max real values are 2088 and 2520 seconds).
Drizzle output resolution changed from 0.03 to 0.05 arcsec, matching the SLACS outputs we've been able to get so far. I kept the dither pattern unchanged at 0.5 pixel offsets, but we should confirm this. (If they chose a dither pattern optimized for supersampling, why wasn't drizzle set to supersample?)
PSF changed from empirical WFC3 to tinytim ACS, since SLACS images were taken . Empirical ACS PSFs surely exist, but I have not found a location where they are available for download.
Lens light added as a double Sersic from (a selection of) the CosmoDC2 catalog. Specifically, we assume the redshift is sufficiently randomized in the simulation, then use the Sersic parameters from the closest (well, searchsorted) entry in the catalog. The selection applied to CosmoDC2 is:
Central galaxy of the halo
Halo mass between 10^13.2 and 10^14.2.
Apparent magnitude (in the i band as defined by LSST) is below 18
This produces a set of lens galaxies that very roughly matches the SLACS lenses in a (redshift, magnitude) plot:
We may want to apply a more sophisticated selection function in apparent magnitude rather than a dumb cutoff at 18. However, I think this is really just a placeholder model before we switch to a parametrized source model we can use in inference.
[This is a draft. I originally did this work on a branch that included some other changes; I need to check if isolating just these changes still results in a working configuration. I should also reproduce and describe the 'early exhaustion' bug that was fixed here.]
This adds new simulation configurations that more closely match the SLACS HST images. Sebastian started this work, the first commit of this PR is just me copying his files into the repo.
The new configs are in the
acs
folder. Besides the mainconfig_train.py
andconfig_val.py
, I added three pairs of configs that gradually relax the assumptions back to those of the paltas paper:noll
represents no lens lightnoll_fixz
also fixes the source and lens redshiftsfantasy
improves the exposure time and resolutionChanges with respect to the paltas paper configuration:
Time per exposure changed from 1320 seconds to Norm(mean=2184/4, std=116/4) seconds. Note we still assume four individual exposures, so this distribution is sampled once, and the total integrated exposure time is four times what we draw from this. This then matches the distribution (2184 +- 116 sec) of SLACS exposure times for the full-orbit observations (at least in mean and std; the min and max real values are 2088 and 2520 seconds).
Drizzle output resolution changed from 0.03 to 0.05 arcsec, matching the SLACS outputs we've been able to get so far. I kept the dither pattern unchanged at 0.5 pixel offsets, but we should confirm this. (If they chose a dither pattern optimized for supersampling, why wasn't drizzle set to supersample?)
PSF changed from empirical WFC3 to tinytim ACS, since SLACS images were taken . Empirical ACS PSFs surely exist, but I have not found a location where they are available for download.
Lens light added as a double Sersic from (a selection of) the CosmoDC2 catalog. Specifically, we assume the redshift is sufficiently randomized in the simulation, then use the Sersic parameters from the closest (well, searchsorted) entry in the catalog. The selection applied to CosmoDC2 is:
This produces a set of lens galaxies that very roughly matches the SLACS lenses in a (redshift, magnitude) plot:
We may want to apply a more sophisticated selection function in apparent magnitude rather than a dumb cutoff at 18. However, I think this is really just a placeholder model before we switch to a parametrized source model we can use in inference.
[This is a draft. I originally did this work on a branch that included some other changes; I need to check if isolating just these changes still results in a working configuration. I should also reproduce and describe the 'early exhaustion' bug that was fixed here.]