• Write a Jupyter Notebook to determine the Tikhonov parameter using a subset of the dataset.

Goal : add the shape constraint used in SCORE to Tikhonet.

• check how the constraint was added in Florent's code.
• add the constraint.
• run a training in an interactive session.

• global context
• description of the problem
• current state of the art
• plan of the paper

We should add a file for all requirements, so that they can be easilty installed.
So far I've had to install

• tensorflow
• scikit-image
• cython
• pywavelets

• adapt Utsav's code from https://github.com/utsav-akhaury/ADMM/tree/main/ADMM_PnP_Radio
• use the Tikhonov filtered image as initialization for the solution (first guess)
• add shape constraint to loss function

• radio
• optical

We need a script to analysis the results. Fadi is working on it here: https://github.com/CosmoStat/ShapeDeconv/blob/master

• load trained model.
• load test dataset.
• evaluate error metrics.
• generate error vs SNR.

• Radio Dataset (MeerKAT3600)
• Optical Dataset

Bug :
In the Unet evaluation notebook, we can observe the presence of anomalies in windows estimation.

Technical explanation :
This due to an unsuccessful output in the function galsim.hsm.FindAdaptiveMom in get_moments. It is flagged with the boolean moments_status of the ShapeData Galsim object.

Physical explanation :
galsim.hsm.FindAdaptiveMom fails to estimate the shape of small sources. Since we're using the pixel resolution of CFHT for HST galaxies, many of these of these galaxies will be very small thus the failure of their shape and window estimation.

Proposed solution :
An easy solution is to remove the outliers (galaxies with failed window estimation).

• remove outliers.
• investigate the statistics of the outliers and the other samples.

Goal : establish a comparison between Tikhonet vs Tikhonet+Shape Constraint vs Sparse Deconvolution vs Sparse Deconvolution+Shape Constraint (SCORE)

• extract a data set for the test.
• run tests on each method.
• set metrics.
• analyse results.

We need a script to train the model. Fadi is working on it here: https://github.com/CosmoStat/ShapeDeconv/blob/master/scripts/tikhonet_train.py

Things to do there:

• : Test the script without constraints (in interactive session)
• : Write a job file for sbatch: http://www.idris.fr/jean-zay/gpu/jean-zay-gpu-exec_mono_batch.html
• : Train a model to the end on JZ

We have to select ranges of source RA & DEC that could fit observations of length 1 2 4 8

• constraint: MeerKAT fixed in SA
• selection in delta vs Delta T
• selection of "appropriate" RA (first pick a fixed date, later, deduced correct RA that fit delta & DT)
• weighting of circumpolar sources vs near horizon sources

• Add noise to the input images in the preprocessing function in

  ShapeDeconv/scripts/tikhonet_train.py

  Line 131 in 3230e23

  def pre_proc_unet(dico):

The noise addition should be done according to the section 4.1 of Deep Learning for space-variant deconvolution in galaxy surveys paper. For each image a SNR level in [20,100] is picked and the corresponding noise is applied to the image.

The current HST PSF (which is Gaussian) is too big. We would like to replace it with an optimally smaller PSF that will be used for all the HST images.

Link to a Jupyter Notebook generate a minimal isotropic COSMOS PSF:
https://github.com/ml4astro/galaxy2galaxy/blob/master/notebooks/BuildingCOSMOS_PSF.ipynb

Link to integrate the PSF to a galaxy2galaxy problem using a fits file:
https://github.com/ml4astro/galaxy2galaxy/blob/6d8b20722a5545c8c79a19cb67c6131c061763ed/galaxy2galaxy/data_generators/cosmos.py#L172

• Adapt the first notebook to generate a minimal isotroptic COSMOS PSF for HST

• Compare the COSMOS PSF to the Gaussian PSF and the true HST PSF and check if it's smaller than the Gaussian one.

• perform a grid search over the shape constraint trade-off parameter.
• compare shape measurement results using the parametric dataset.
• compare MSE results using the parametric and/or realistic dataset.

Here are the components we want to extract and make easily accessible:

• Data input tools: everything needed to load data
• Shape constraint losses: we want the shape constraint losses to be standalone functions

• State of the art
• Tikhonet
• Shapenet