Backward-incompatible PSF/linear solver problem?

[Maverick-Oh]

There seems to be a recent (within 1 year?) backward-incompatible change related to PSF and linear solver part. I wasn't able to narrow down what specific update it is, but I compared two versions of lenstronomy--1.11.5 (the most recent version, 'new version' now on) and 1.10.3 ('old version' now on).

I was loading old fitting results, which include chain_list and fitting_seq, to check the result again and possibly change some settings to rerun.

For the same fitting sequence, the old version gave me a reasonable model plots. However, the new version gives weirdly different sets of 'amp' values for all profiles and the residual looks very off, showing the 'squares' of PSF convolution. I think something is very off.

Can I get to know what you think?

To check quickly, please see the PDFs below. It's loading the same file but giving very different 'amp' values and model plots.

fitting_result_old_1.10.3.pdf
fitting_result_new_1.11.5.pdf

To check the notebook, please download the two notebooks and the pickled fitting result file below. You would need to change
my_dir = '../../data2use/20220519_pemd_nomultipole' to the directory that has the pickled file.

fitting_results_ipynbs.zip

Dropbox link for the fitting_seq file to run the notebooks

[sibirrer]

@Maverick-Oh I think it had to do with the change in the psf_error_map, see this pull request: #423

[Maverick-Oh]

@sibirrer Hello Simon, thanks for guiding me right to the spot I need to look into. I was able to solve the issue. Thanks a lot!