Core functionality all seems to work now. Just stuck on why the generate_skymodels step is still executed even though the array length checked in the when condition should be >0.

why the generate_skymodels step is still executed even though the array length checked in the when condition should be >0.

This looks like it's maybe just a Toil oddity.

Functionality all works, but the vlass starting models are not used it seems. A manual facetselfcal run against the same VLASS image yields wildly different astrometry even if I completely remove the generat_skymodels step from the CWL code as a test, so I suspect the pipeline is still generating a point source and passing that.

Left to right: VLASS image, pipeline image 9, manual image 3

It looks like the difference may be in the PyBDSF commands for source finding. The facetselfcal-generated model has more components than what skynet outputs:

Skynet

# (Name, Type, Patch, Ra, Dec, I, Q, U, V, MajorAxis, MinorAxis, Orientation, ReferenceFrequency='144e+06', SpectralIndex='[]', LogarithmicSI) = format
, , P0, 00:00:00, +00.00.00
ME0,GAUSSIAN,P0,10:42:07.07427757,+56.20.35.11342256,0.7501937174536204,0.0,0.0,0.0,1.6648132122494166,1.2344896254335085,109.43104961290301,144e+06,[-0.7],true
ME1,GAUSSIAN,P0,10:42:06.08426152,+56.20.42.68432263,0.0933097437405896,0.0,0.0,0.0,1.6331785957898932,0.34046779089727963,136.72680994037083,144e+06,[-0.7],true

Facetselfcal

format = Name, Type, Patch, Ra, Dec, I, Q, U, V, MajorAxis, MinorAxis, Orientation, ReferenceFrequency='3.00000e+09', SpectralIndex='[]'

, , ILTJ104206.59+562039.5_vlass_patch_s0, 00:00:00, +00.00.00
ILTJ104206.59+562039.5_vlass_w0_i0_s0_g0, POINT, ILTJ104206.59+562039.5_vlass_patch_s0, 10:42:7.119336, +56.20.35.033496, 2.555e-02, 0.0, 0.0, 0.0, 0.00000e+00, 0.00000e+00, 0.00000e+00, 3.00000e+09, [-0.8]
ILTJ104206.59+562039.5_vlass_w0_i0_s0_g2, POINT, ILTJ104206.59+562039.5_vlass_patch_s0, 10:42:6.991234, +56.20.35.438892, 1.120e-02, 0.0, 0.0, 0.0, 0.00000e+00, 0.00000e+00, 0.00000e+00, 3.00000e+09, [-0.8]
ILTJ104206.59+562039.5_vlass_w0_i0_s0_g3, GAUSSIAN, ILTJ104206.59+562039.5_vlass_patch_s0, 10:42:6.768397, +56.20.35.979193, 6.761e-03, 0.0, 0.0, 0.0, 4.41003e+00, 0.00000e+00, 1.68543e+02, 3.00000e+09, [-0.8]
, , ILTJ104206.59+562039.5_vlass_patch_s1, 00:00:00, +00.00.00
ILTJ104206.59+562039.5_vlass_w0_i0_s1_g1, POINT, ILTJ104206.59+562039.5_vlass_patch_s1, 10:42:6.086021, +56.20.42.687386, 3.819e-03, 0.0, 0.0, 0.0, 0.00000e+00, 0.00000e+00, 0.00000e+00, 3.00000e+09, [-0.8]
, , ILTJ104206.59+562039.5_vlass_patch_s2, 00:00:00, +00.00.00
ILTJ104206.59+562039.5_vlass_w0_i1_s2_g4, GAUSSIAN, ILTJ104206.59+562039.5_vlass_patch_s2, 10:42:4.669273, +56.21.1.664814, 2.484e-03, 0.0, 0.0, 0.0, 2.34113e+00, 9.84720e-01, 6.75894e+01, 3.00000e+09, [-0.8]

After some digging around in skynet.py it is indeed related to how it manipulates the sky model found by PyBDSF. Writing out the PyBDSF model directly results in the same model as what facetselfcal generates.

This means as far as I can tell the goal of this branch is there, so I'll mark it ready for review and we can fix skynet in another one.

This looks like it's maybe just a Toil oddity.

@tikk3r could you create an issue for this on our issue tracker? I can look into it later.

Alright, I finally figured out what is going wrong in this branch. I still think there are issues with how skynet manipulates the model, but that is not the main cause of the selfcal problem here. For some reason cwltool (or at least my version 3.1.20250110105449) validates the workflow despite there being a logic error in it. The model is generated via this call:

    - id: generate_skymodels
      in:
        - id: msin
          source: msin
          valueFrom: $(self[0])
        - id: delay_calibrator
          source: delay_calibrator
        - id: process_all
          default: true
        - id: starting_skymodel
          default: starting_skymodel
      out: 
        - id: skymodel
        - id: logfile
      run: ../../steps/delay_cal_model.cwl
      when: $(inputs.starting_skymodel == null)

however, delay_cal_model.cwl does not have a starting_skymodel input ; only model_image. It doesn't catch this apparently.

The issue with wrong skymodels seems to be resolved after remove the unnecessary step from the subworkflow. Here's an updated comparison between the selfcal and the VLASS model, now showing good agreement in astrometry as would be expected.

This looks like it's maybe just a Toil oddity.

@tikk3r could you create an issue for this on our issue tracker? I can look into it later.

Do we still want this now that you've found a Toil ticket and opened a PR about it?