Skip to content

feat: prior configs + docs for the AdaptPower regularization siblings - #592

Merged
Jammy2211 merged 1 commit into
mainfrom
feature/adapt-linear-regularization
Aug 29, 2026
Merged

Jammy2211 merged 1 commit into
mainfrom
feature/adapt-linear-regularization

Conversation

@Jammy2211

@Jammy2211 Jammy2211 commented Aug 29, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

Prior configs, API docs and a composition test for the four *Power regularization siblings added in PyAutoLabs/PyAutoArray#512 (AdaptPower, AdaptSplitPower, AdaptSplitZerothPower, MaternAdaptPowerKernel).

  • autogalaxy/config/priors/regularization/{adapt_power,adapt_split_power,adapt_split_zeroth_power,matern_adapt_power_kernel}.yaml are copies of their legacy counterparts — same LogUniform(1e-6, 1e6) coefficient priors, which under the new classes span λ² rather than λ⁴ — plus power: {type: Constant, value: 1.0} so the convention switch is never sampled by a search.
  • docs/api/pixelization.rst gains AdaptPower / AdaptSplitPower in the ag.reg autosummary.
  • test_autogalaxy/test_regularization_power.py proves the ag.reg re-export chain, that af.Model exposes exactly inner_coefficient / outer_coefficient / signal_scale (not power), and that the identifier differs from the legacy class.

CI note: green here already. The PyAutoHeart reusable workflow clones each dependency library at the matching branch if it exists, and PyAutoArray#512 uses the same branch name, so this PR's CI tested against it. Library-first merge order still applies for main's sake: merge PyAutoLabs/PyAutoArray#512 first, otherwise main here would reference ag.reg.AdaptPower before the symbol exists.

API Changes

None to autogalaxy source — config, docs and test additions only. The new public symbols are added upstream in PyAutoLabs/PyAutoArray#512 and reach users here as ag.reg.AdaptPower, ag.reg.AdaptSplitPower, ag.reg.AdaptSplitZerothPower, ag.reg.MaternAdaptPowerKernel.
See full details below.

Test Plan

  • pytest test_autogalaxy/ — 1152 passed (with PyAutoArray#512 on PYTHONPATH)
  • af.Model(ag.reg.AdaptSplitPower) composes; power resolves to the Constant value 1.0 and is absent from the prior tuples
  • af.Model(ag.reg.AdaptSplit).identifier != af.Model(ag.reg.AdaptSplitPower).identifier
  • CI green on all four legs (unittest 3.12 / 3.13 / nojax, docs-build), tested against the PyAutoArray branch
Full API Changes (for automation & release notes)

Added

  • Prior configs for AdaptPower, AdaptSplitPower, AdaptSplitZerothPower and MaternAdaptPowerKernel — the coefficient priors are unchanged copies of the legacy schemes' (LogUniform(1e-6, 1e6)), and power is a Constant prior fixed at 1.0.
  • docs/api/pixelization.rst autosummary entries for AdaptPower and AdaptSplitPower.

Migration

  • Before: ag.reg.AdaptSplit(inner_coefficient=c, outer_coefficient=c)
  • After: ag.reg.AdaptSplitPower(inner_coefficient=c**2, outer_coefficient=c**2)
  • The legacy schemes and their prior configs are untouched, so no migration is required — this is guidance for new work.

Depends on PyAutoLabs/PyAutoArray#512. Refs PyAutoLabs/PyAutoArray#511.

Generated by the PyAutoLabs agent workflow.

Adds prior configs for the four *Power regularization classes introduced in
PyAutoArray (PyAutoLabs/PyAutoArray#512): adapt_power, adapt_split_power,
adapt_split_zeroth_power and matern_adapt_power_kernel. Each is a copy of its
legacy counterpart — the same LogUniform(1e-6, 1e6) coefficient priors, which
now span lambda^2 rather than lambda^4 — plus a Constant prior fixing power at
1.0 so a non-linear search never samples the convention switch.

Also adds AdaptPower / AdaptSplitPower to the docs/api pixelization
autosummary, and a composition test proving the ag.reg re-export chain and the
prior-config resolution.

Depends on PyAutoLabs/PyAutoArray#512. Refs PyAutoLabs/PyAutoArray#511
@Jammy2211 Jammy2211 added the pending-release PR queued for the next release build label Aug 29, 2026
@Jammy2211
Jammy2211 merged commit 2ee44d5 into main Aug 29, 2026
4 checks passed
@Jammy2211
Jammy2211 deleted the feature/adapt-linear-regularization branch August 29, 2026 20:33
@Jammy2211 Jammy2211 removed the pending-release PR queued for the next release build label Sep 4, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant