Result Analysis

[fadinammour]

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

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

[fadinammour]

The inputs and the targets in the dataset seem to be the same.

[fadinammour]

Subtasks progression

• [x] Evaluating SCORE metrics was straight forward because the functions were implemented using numpy.

• [x] Implementing the KSB metric (the same one used in the Tikhonet paper) is more challenging because it uses a binary from the shapelens package that was modified in a private repository. In addition, the binary is meant to be used with command lines.

• [x] Estimating the SNR, how should it be done on HST COSMOS data for it to be satisfying according to optical astrophysics standards?

[fadinammour]

Instead of evaluating the SNR, the magnitude was used since it is available in the data. KSB had many failures on the realistic galaxies dataset so it was replaced with the adpative moments (see FindAdaptiveMoments in Galsim HSM).