/RAIL

Redshift Assessment Infrastructure Layers

Primary LanguagePythonMIT LicenseMIT

RAIL: Redshift Assessment Infrastructure Layers

This repo is home to a series of LSST-DESC projects aiming to quantify the impact of imperfect prior information on probabilistic redshift estimation. RAIL differs from PZIncomplete in that it is broken into stages, each corresponding to a manageable unit of infrastructure advancement, a specific question, and a potential publication opportunity. By pursuing the piecemeal development of RAIL, we aim to achieve the broad goals of PZIncomplete.

Contributing

The RAIL repository uses an issue-branch-review workflow. When you identify something that should be done, make an issue for it. To contribute, isolate an issue to work on and leave a comment on the issue's discussion page to let others know you're working on it. Then, make a branch with a name of the form issue/#/brief-description and do the work on the branch. When you're ready to merge your branch into the master branch, make a pull request and request that other collaborators review it. Once the changes have been approved, you can merge and squash the pull request.

Immediate Plans

An outline of the baseline RAIL is illustrated here.

  1. MonoRAIL: Build the basic infrastructure for controlled experiments of forward-modeled photo-z posteriors
  • a rail.creation submodule that can generate true photo-z posteriors and mock photometry
  • an rail.estimation submodule with a class for photo-z posterior estimation routines, including a template example implementing the trainZ (experimental control) algorithm
  • an rail.evaluation.metric submodules that calculate the metrics from the PZ DC1 Paper for estimated photo-z posteriors relative to the true photo-z posteriors
  • documented scripts that demonstrate the use of RAIL in a DC1-like experiment on NERSC
  • an LSST-DESC Note presenting the RAIL infrastructure
  1. RAILroad: Quantify the impact of nonrepresentativity (imbalance and incompleteness) of a training set on estimated photo-z posteriors by multiple machine learning methods
  • a rail.creation.degradation submodule that introduces an imperfect prior of the form of nonrepresentativity into the observed photometry
  • at least two rail.estimation.estimator wrapped machine learning-based codes for estimating photo-z posteriors
  • additional rail.evaluation.metric modules implementing the qp metrics
  • documented scripts that demonstrate the use of RAIL in a blinded experiment on NERSC
  • an LSST-DESC paper presenting the results of a controlled experiment of non-representativity

Future Plans

The next stages (tentative project codenames subject to change) can be executed in any order or even simultaneously and may be broken into smaller pieces each corresponding to an LSST-DESC Note.

  • Extend the imperfect prior models and experimental design to accommodate template-fitting codes (name TBD)
  • Off the RAILs: Investigate the effects of erroneous spectroscopic redshifts (or uncertain narrow-band photo-zs) in a training set
  • Third RAIL: Investigate the effects of imperfect deblending on estimated photo-z posteriors
  • RAIL gauge: Investigate the impact of measurement errors (PSF, aperture photometry, flux calibration, etc.) on estimated photo-z posteriors
  • DERAIL: Propagate the impact of imperfect prior information to 3x2pt cosmological parameter constraints
  • RAIL line: Implement a more sophisticated true photo-z posterior model with SEDs and emission lines