PSF rewrite exploration checklist

[benmontet]

Hi everyone,

Here is the checklist of things we would like to do/explore. Please do suggest other things and I'll add them to the list! This is at present copied from our notes on slack.

• Can we use more accurate PRF models?
  • How do we build a fast analytic model of the 'true' PSF that can be fit at each cadence?
  • Add in that model into the 'models' class and let that be called with the current setup
  • Verify it does a good job
• Can we make better use of autodifferentiation?
  • Does this mean a full rewrite in tensorflow2 with adam? Is that fast?
  • Would alternatives like pytorch work just as well, or is tensorflow2 the best bet?
  • Can we build a system that uses not tf1.x that does as good a job as quickly with the current simple models we have?
• Can we leverage all PSF models to make better photometry?
  • How nearby a star can we use the psf results for as a 'regressor' to improve the photometry?
  • Do we win a lot by combining information from many stars together? Or is there other telescope information we can include?
  • Do we want to use the Guassian parameters here, or can some other type of PSF model parameters do much better?
  • Can we build a function that does these bits automatically?