Connecting BLISS and JIF

[ismael-mendoza]

Connecting BLISS and JIF

We will leverage existing DESC tools to build a bayesian pipeline for obtaining shear posteriors from simulated images.

Contacts: @ismael-mendoza Day/Time: Afternoons during the Sprint week Main communication channel: #desc-bayesian-pipelines GitHub repo:

• https://github.com/LSSTDESC/bayesian-pipelines-pixels
• https://github.com/prob-ml/bliss
• https://github.com/mdschneider/JIF Zoom room (if applicable): TBD

Goals and deliverable

Our goal is to get a proof-of-concept pipeline that goes from images to weak lensing shear posteriors. The concrete deliverable will be a notebook that demonstrates obtaining such shear posteriors leveraging two existing tools: BLISS and JIF.

Resources and skills needed

Anyone with an interest in Bayesian statistics, probabilistic cataloging, and/or weak lensing is encouraged to join. Some knowledge of python would be helpful. We will start with some short tutorials on running BLISS and JiF so anyone can get up to speed.

Detailed description

Here is an outline for the steps we will follow (rougly ~ per day of sprint week):

1. Tutorials on BLISS and JIF + current structure of forward/generative model on pixel pipeline
2. Ensure generative model of images using galsim in bpp are suitable. Use the HDF5 format to save them into disk. Add dataset/dataloader class in BLISS that can train on these images. Train BLISS on these saved images (overnight).
3. Run BLISS on test dataset and save samples per image into HDF5 format. Ingest samples from BLISS into JIF
4. Analyze results, validate shear posteriors from JIF
5. Repeat process and validate with descwl-shear-sims repo.

Description of tools needed

BLISS - Is a fully probabilistic deblender that can output probabilistic catalogs of counts, locations, fluxes given an image. It can also output samples of deblended images that capture uncertainty in location and degree of blending. BLISS will be used to posterior samples of counts, locations, and deblended images to the shear inference algorithms.

JIF - This algorithm fits a parametric model to a single galaxy and MCMC the parameter space, propagating to shear. The project will test this as a viable tool for shear inference when coupled with BLISS, which is expected to provide JIF with posterior samples of counts and locations.

[arunkannawadi]

Will the tutorial be recorded?

[ismael-mendoza]

Yes we will definitely record it :)

[arunkannawadi]

Is anything happening this morning, before the tutorial?

Edit: I got the answer from Ismael, and nothing is happening.