jballet
commented
1 year ago I have managed to reproduce the failure on a much smaller file, by selecting only 1 to 10 GeV and a 200x200 piece of the source map (extracted from the previous source map, so the projection center is not at the center of the map). The test harness is available at this link. In that case all other sources and the isotropic are completely fixed and included into the FixedSources component. You will see by comparing with the PNG files in the previous example that the source is not found exactly at the same place. The circle is larger and much closer to what it should be (when it works). The failure is still sensitive to the exact pixellization, because when I recreate the map from scratch (with projection center at the center of the map) it does not fail any more (and the best fit position is again at a different place, although all the positions are within the 68% error circle).
ndilalla
I am trying to routinely localize all catalog sources in GAL coordinates. In about 10% of my calls I get the warning message 2023-08-03 18:44:30 WARNING GTAnalysis._localize(): Fit to localization contour failed. 2023-08-03 18:44:30 WARNING GTAnalysis._localize(): Localization failed. Keeping existing position. and the fit_success entry turns to False, so my run stops. It appears that fermipy did find a best localization (it even provides it in the log) but, as the warning says, failed to extract an error ellipse from the TS contours (and indeed, the error plotted on the localization plot looks wrong). This is the first (large-scale) localization map (called _localize).
This is the close-up (called _localize_peak) that I think illustrates the problem.
I suspect the algorithm to fit the ellipse is not robust enough.
This warning is sometimes (not always) accompanied by
fermi_py_patch514/lib/python3.9/site-packages/scipy/optimize/
_lsq/common.py:147: RuntimeWarning: invalid value encountered in double_scalars
alpha = max(0.001 alpha_upper, (alpha_lower alpha_upper)**0.5)
I should add that I am leaving a lot of freedom in the neighboring sources at the localization step (because it is in theory better to do so). But it does not seem to be the problem because the fit itself does not fail and the TS map looks smooth.
The failure is very sensitive to details. If I fix a neighboring source it will often go away.
Consequently I have not been able to streamline the input as much as I normally do and the input srcmap file is quite big (although the computing time is only 3 minutes or so).
The test harness can be found at this link until August 18.
Call fermipy_1073.py.
The big files (gll_iem_v07.fits and ft2_12years_reduced.fits) are not included. The former is the regular diffuse model. The latter is available at SLAC at /nfs/farm/g/glast/g/catalog/P8_P305.
The products of my run (PNG files and log) are included in the tgz file.
The fermipy version I use is the patch514 that was created to fix issue 514. It is called 1.2.0+2.gfa69 by conda list. It is associated with FermiTools 2.2.0.