Roadmap: PSF Homogenization feature

[ismael-mendoza]

Roadmap for issue #378. This is intended to layout a set of concrete tasks/issues for @wangchv to work on. @jeff-regier let me know if you would add anything or have any suggestions.

Simulated

• create a full image decoder that only uses galsim
• create a PsfSampler class that returns a random sample of psfs from a file or galsim
  • Create functions like .from_file, .from_array and a subclass that uses galsim with varying FWHM
• extend full image decoder to output PSF homogenized images using functions in homogenization.py
• train detection_encoder on homogenized dataset that uses a random galsim PSF, check results are reasonable
• compare performance between encoders trained on homogenized images vs not (precision, recall, etc.)

SDSS Data

• Reimplement psf_at_points in sdss.py. See this diff.
• Decide on a set of frame/camcols/runs to use in SDSS that overlap with DeCals ?
  • Might need to close or make progress in use DECaLS for ground truth instead of SDSS coadd #623
• Use psf_at_points to create a sample dataset of PSFs across previously decided frame/camcols/runs/position
• train encoders on homogenized dataset using random sample of these real PSFs
  • the psfs extracted in the previous tasks should be used as input to PsfSampler class
• evaluate performance of encoders trained on real PSF homogenized dataset across multiple frames vs one trained on a single PSF

If time

• Train and evaluate binary_encoder and galaxy_encoder as well

[jeff-regier]

Great!

[jeff-regier]

It's been difficult to homogenize the images adequately -- let's try encoding the PSF instead (#694)