Demonstrate Model Bias with real galaxies

b-remy commented 2 years ago

This pull request is to document the implementation of a script for the joint inference of shear and shape for #5 .

So far, we assume:

Gaussian profiles (fixed hlr)
COSMOS PSF (fixed)
Constant shear (varying)
Intrinsic ellipticity (varying)

Remarks:

tf.function was enable by removing all non-tensorflow code in the edward model
Increasing the number of galaxies seems to unbias the shear estimation

b-remy commented 2 years ago

Also sometimes it infers -gamma_1 insted of gamma_1.

b-remy commented 2 years ago

The last version of the code is working well as illustrated in #5 .

I'll try to reproduce the issue I had above. In this scenario there were to calls to the model, which may trouble the code idk...

with ed.tape() as true_params:
    ims = model(N*N, stamp_size)

# Apply a constant shear on the field
custom_shear = [0.01, 0.]
with ed.condition(hlr=true_params['hlr'],
                  gamma=custom_shear,
                  e=true_params['e'],
                  ):
    ims = model(N*N, stamp_size)

b-remy commented 2 years ago

This pull request contains a lot of code a files, but can be summarized as follow:

refactor the code as a package
gems/ed_utils.py enables batch computation through edward likelihoods
gems/models.py contains all the probabilistic models that I implemented so far (we may want to split this file into pieces in the future)
gems/psf.py is a utils file to get some psf models (not used in the latest PGM where we provide the PSF as input to the model, but needed for other scripts as scripts/simpler_model.py)
gems/shear.py contains code modeling the shear at the map level assuming a GRF prior
gems/utils.py utils code, mostly for the shear map model
scripts/cosmo_shear_paramtric.py and scripts/cosmo_shear_real.py are scripts running inference on PGM using sersic light profiles to fit observations with 1) real galaxies 2) parametric galaxies
scripts/simpler_model.py script to demonstrate model bias on observations generated with light profile only (e.g. gaussians on gaussians, gaussians on sersics, ...)
scripts/varying_shear_gaussians.py is a script to perform inference using a PGM with a shear GRF prior

Merging this pull request would close the following issues:

1 which describes how we set prior models for light profiles galaxies based on the COSMOS catalog sersic fit
5 where we wanted to demonstrate that we could recover unbiased contours of the shear using gaussian light profile on gaussian galaxies observation
11 describing experiment results fitting parametric models to both real and parametric galaxies, demonstrating model bias on real galaxy observations

b-remy / gems

Demonstrate Model Bias with real galaxies #7

1 which describes how we set prior models for light profiles galaxies based on the COSMOS catalog sersic fit

5 where we wanted to demonstrate that we could recover unbiased contours of the shear using gaussian light profile on gaussian galaxies observation

11 describing experiment results fitting parametric models to both real and parametric galaxies, demonstrating model bias on real galaxy observations