Skip to content

rectangular_adapt_constant_split_guard.md

Latest commit

 

History

History
137 lines (107 loc) · 4.76 KB

rectangular_adapt_constant_split_guard.md

File metadata and controls

137 lines (107 loc) · 4.76 KB
Raw

Users keep combining RectangularAdaptDensity meshes with ConstantSplit

Type: feature Target: PyAutoArray Difficulty: too-large Autonomy: supervised Priority: high Status: formalised

Users keep combining RectangularAdaptDensity meshes with ConstantSplit regularization in pixelized source models, for example:

pixelization = af.Model(
    al.Pixelization,
    mesh=af.Model(al.mesh.RectangularAdaptDensity, shape=(28, 28)),
    regularization=al.reg.ConstantSplit,
)

This must be flagged as an error before the fit runs. The combination is currently invalid and can produce a pathological all-zero pixelized source in the output products instead of failing clearly.

Concrete repro:

  • Data/script context: @z_help/jacob/HerBS-28.ms.shifted_fits/jax_modelling_template.py
  • Verbatim capped reproduction: @z_help/jacob/HerBS-28.ms.shifted_fits/jax_modelling_template_verbatim_nlike1000.py
  • Pixelization output path from the repro: output/interferometer/verbatim_nlike1000/pixelization/01c465702e769ad81c9c0e35b3d232d0

Observed behavior from that run:

source_plane_images.fits HDU 0: sum_abs=0.0 nonzero=0
source_plane_images.fits HDU 1: sum_abs=0.0 nonzero=0
fit_dirty_images.fits DIRTY_MODEL_IMAGE: sum_abs=0.0 nonzero=0
Maximum Log Likelihood = -12395933.858

Rebuilding the reported max-likelihood fit eagerly and accessing the inversion reconstruction crashes in the split regularization path:

IndexError: index 4 is out of bounds for axis 0 with size 4

The exception occurs in:

@PyAutoArray/autoarray/inversion/regularization/regularization_util.py
reg_split_np_from()

called from:

@PyAutoArray/autoarray/inversion/regularization/constant_split.py
ConstantSplit.regularization_matrix_from()

Root cause:

ConstantSplit assumes split-point interpolation rows have enough padded mapping slots to append the parent pixel when the parent pixel is not already present. With RectangularAdaptDensity, the split mappings can have only four slots. If all four are already occupied and the parent pixel is absent, reg_split_np_from() tries to write to slot j + 1 == 4, which is out of bounds for an array with size 4. In JAX/vectorized modeling this can surface as a silently bad inversion / zero source output rather than a clear exception.

Desired fix:

Add an explicit validation guard which rejects this unsupported configuration early, with an error message that tells users what to do instead.

The guard should catch both concrete and model-composition forms:

al.Pixelization(
    mesh=al.mesh.RectangularAdaptDensity(...),
    regularization=al.reg.ConstantSplit(...),
)

and:

af.Model(
    al.Pixelization,
    mesh=af.Model(al.mesh.RectangularAdaptDensity, ...),
    regularization=al.reg.ConstantSplit,
)

Suggested error text:

ConstantSplit regularization is not supported with RectangularAdaptDensity
meshes. This combination can produce invalid split regularization stencils and
all-zero pixelized source outputs. Use al.reg.Constant with
RectangularAdaptDensity, or use ConstantSplit with a Delaunay mesh.

Likely implementation locations to assess:

  • @PyAutoArray/autoarray/inversion/pixelization.py if concrete Pixelization construction can validate mesh / regularization types.
  • The inversion / mapper construction path if concrete construction is too early or incomplete.
  • The AutoFit model-analysis pre-fit validation path may be needed to catch af.Model(al.Pixelization, ...) before sampling begins. If the validation naturally belongs outside PyAutoArray for af.Model objects, add the concrete PyAutoArray guard first and add a companion AutoLens / analysis guard where the prior model can be inspected.

Verification:

  1. Add unit tests for concrete Pixelization construction or first use: RectangularAdaptDensity + ConstantSplit must raise the clear error.
  2. Add a model-composition test, or an analysis pre-fit test, showing that af.Model(al.Pixelization, mesh=af.Model(al.mesh.RectangularAdaptDensity), regularization=al.reg.ConstantSplit) fails before Nautilus starts.
  3. Confirm allowed combinations still work:
    • RectangularAdaptDensity + Constant
    • Delaunay + ConstantSplit
  4. If touching the split regularization utility itself, add a low-level regression test for a split mapping row where the parent pixel is absent and all four slots are occupied, so the code cannot silently write out of bounds or produce invalid stencils.

Do not paper over this by clipping indices or returning an all-zero reconstruction. This should be a hard, user-facing configuration error until the split-regularization stencil code is redesigned to support rectangular adaptive meshes.