(Old version?) PyBDSM takes forever (or hangs) inside Stimela container

[o-smirnov]

Got the following simple recipe courtesy of @Athanaseus:

import stimela
import sys

INPUT = '.'
OUTPUT = '.'
MSDIR = 'msdir'

IMAGE = sys.argv[1]
PIX_THRESH = int(sys.argv[2])
ISL_THRESH = int(sys.argv[3])


stimela.register_globals()

recipe = stimela.Recipe('Source Finder', ms_dir=MSDIR)
#JOB_TYPE='singularity', singularity_image_dir=os.environ["STIMELA_PULLFOLDER"])


recipe.add('cab/pybdsm', 'source_finder',
        {
            "filename"        :   "{}:output".format(IMAGE),
            "outfile"            :   "pybdsm_{}".format(IMAGE[:-5]),
            "thresh_pix"     :    PIX_THRESH,
            "thresh_isl"      :    ISL_THRESH,
            "clobber"          :    True,
            "catalog_type" :    "gaul",
            "group_tol"    : 12,
            "adaptive_rms_box" : True
        },
            input=INPUT,
            output=OUTPUT,
            label='src_finder:: Sourcery')

recipe.run()

Funning it in /net/young//home/oms/projects/OldDevils/selfcal-4C12.03/test with 1558752655-MFS-image.fits and thesholds of 50, 30, the thing loops forever late in the Gaussian fitting stage:

# Fitting islands with Gaussians .......... : [-] 357/390

There's one CPU core using 100%. Some of these runs eventually succeed (after e.g. 18 hours!!)

Running the same job natively in the KERN version of pybdsf, with the same options, the job finishes in <10mins. Curiously though, it only fits 319 Gaussians.

So I assume there's something different in the Stimela defaults compared to the normal PyBDSM defaults. But what? And if so, this is misleading and should be fixed.

[KshitijT]

Curiously though, it only fits 319 Gaussians.

Possibly some of the Gaussians were flagged in the fitting ?

[o-smirnov]

But why the different flagging behaviour inside/outside of Stimela then?

[KshitijT]

But why the different flagging behaviour inside/outside of Stimela then?

Actually, what's the number of Gaussians fit in the other case? Is it significantly different? (I am suspecting version differences).

[o-smirnov]

No sorry, the number of Gaussians is a red herring, I had my thresholds set backwards. When I set them consistently with the recipe above, I get 390 Gaussians in both cases. Native version still finishes in reasonable time.

I suspect the Stimela version is super old, I can see the Stimela base is 1.2.0. I tried to update this to 1.6.0, but now the container fails with

# ModuleNotFoundError: No module named 'bdsf'

I think this is because PyBDSM needs to be run with Python 2.7 still, while the new Stimela base uses Python 3 by default. There's a way to switch it to 2.7, but I need @SpheMakh to tell us how.

[o-smirnov]

@SpheMakh bump. This is holding us up badly. Please either roll us a new PyBDSM container, or teach a man to fish, i.e. remind me how to make a Stimela container where run.py is forced to run in python 2.7 instead of 3.x.