TODOs

3/11/15 Update (meeting w/ Albert): Toward finishing the Celeste model and scaling it up, we have three types of tasks

• Model (generating model images from)

  • Galaxy implementation w/ Dustin's model (Andy)
  • Quasar from pixels w/ ICML QSO model (Andy)
  • Switching between types (Gal, Star, QSO)
  • Gal, Star, QSO Priors (structured prior - nonparametric mixtures)

• Inference:

  • Sample source types (a_s = {Gal, Star, QSO})
  • Sample photon origins, and conditioned on:
    • Galaxy (sample transformation, size, etc)
    • QSO (sample basis weights and z)
    • Star (sample temperature, lum/dist)
      • Need to adopt something more than univariate slice sampling (HMC/Tempering)
  • Implement adaptive patches (for a 2 MPX image, we only need to draw a line around each source)
  • Parallelization?
    • Where can this be parallelized? What is the best way to do
  • Reversible Jump (For now, we might just stick with a good initialization and worry about this later)

• Evaluation/Data Munging

  • Collect Gal, Star, QSO patches for early testing and classification
  • Optimize code for NERSC
  • Do Stripe 82 test (re-create Jeff's paper)
  • Hubble Truth Table test (3 million objects w/ information uncorrupted by the atmosphere - David mentioned this in a meeting)
  • Generate a catalog from scratch from a 2 MPX image
    • estimate how much time the entire catalog will go through all SDSS images

10/7/14 Update:

• Robust regression fit for spectrum curves. Examine structure in residuals. (Andy)
  • Think about clustering models for types of stars based on spectrum (toward an informative prior that helps identify luminosity and distance).
• Clean up birth/death + RJMCMC moves and interface w/ celeste model. (Albert)
• Implement EM to optimize over temperature/brightness variable for starting point/sanity check. (Andy)

9/23/14 Update:

• Slice sample brightness + locations for fixed-num-source stamps (Albert)
• Incorporate temperature function into likelihood (Andy)
• Validate samplers on synthetic data (ground truth params) (Andy + Albert)
  • Examine ESS, Autocorr, R Hat (Andy)
• Optimize Gaussian Mixture Model computation

9/16/14 Update:

• Compute and save a handful of images, including pixel counts and metadata (image geometry variables). Verify that the counts correspond to the ones that enter into the likelihood. (Albert)

• Generate synthetic data from the latest graphical model (in Dropbox/rnd.astro/celeste/celeste_2014-09-08.pdf) and visualize. Figure out which parts are handled by tractor and which need to be implemented. (Andy)

  • First just generate stars (using tractor model images + poisson noise)

• Compute likelihood function over u, t, b, $\theta$. (Andy)

  • for stars:
    • examine/plot profile likelihoods for individual (u_ra, u_dec) and five brightnesses (using tractor model images + poisson likelihood)
    • implement/re-purpose code for non-dimension changing MCMC for these parameters.
      • assess validity (guided by profile likelihood and ground truth parameters)
      • run chain diagnostics

• Schedule a regular meeting for the fall.

  • Andy/Albert Weekly 2 hour block (+ RPA every other week ... bi-weekly?)

• Import important bits of Brenton's code into new repo.

• Revisit questions about units, document in wiki.

• Get to where we can evaluate likelihood by rendering an image.

• Get stars working.

• Get galaxies working.