perf: cache immutable @property on Fit classes

[Jammy2211]

Overview

37 @property methods across FitDataset, FitImaging, and FitInterferometer recompute on every access despite depending only on immutable constructor args. The worst case is FitImaging.model_data which rebuilds the entire inversion (~5-20s for Delaunay sources) on every access — and residual_map, chi_squared_map, log_likelihood all access it indirectly, creating a cascade of redundant recomputation.

Discovered during quick-update rendering profiling: accessing fit.model_data + fit.subtracted_images_of_planes_list cost 13.7s from two redundant inversion recomputations. Changing these to @cached_property eliminates the cascade.

Plan

• Sweep all @property methods on FitDataset (PyAutoArray), FitImaging / FitInterferometer (PyAutoGalaxy + PyAutoLens) and change expensive ones to @cached_property
• Focus on the recomputation cascades: model_data → residual_map → chi_squared_map → log_likelihood, and blurred_image → profile_subtracted_image → model_data
• Verify no pickling or JAX pytree issues (audit confirms none of these classes use __getstate__ or pytree registration on Fit objects)
• Leave cheap leaf-access properties (data, noise_map, mask, _xp) uncached
• Run full test suites across all three repos + profiling scripts to confirm speedup

Detailed implementation plan

Affected Repositories

• PyAutoArray (primary — FitDataset base class)
• PyAutoGalaxy (FitImaging, FitInterferometer)
• PyAutoLens (FitImaging override)

Branch Survey

Repository	Current Branch	Dirty?
PyAutoArray	main	clean
PyAutoGalaxy	main	CLAUDE.md
PyAutoLens	main	CLAUDE.md, README

Suggested branch: feature/cache-fit-properties

Properties to cache (37 total)

FitDataset (PyAutoArray) — 12 properties:
chi_squared, chi_squared_map, log_likelihood, log_evidence, log_likelihood_with_regularization, normalized_residual_map, noise_normalization, reduced_chi_squared, residual_map, residual_flux_fraction_map, signal_to_noise_map, figure_of_merit, grids

FitImaging (PyAutoGalaxy) — 8 properties:
blurred_image, profile_subtracted_image, model_data, galaxy_image_dict, galaxy_model_image_dict, model_images_of_galaxies_list, subtracted_images_of_galaxies_dict, subtracted_images_of_galaxies_list

FitImaging (PyAutoLens) — 5 properties:
blurred_image, profile_subtracted_image, model_data, subtracted_images_of_galaxies_dict, subtracted_signal_to_noise_maps_of_galaxies_dict, subtracted_images_of_planes_list

FitInterferometer (PyAutoGalaxy) — 6 properties:
profile_visibilities, profile_subtracted_visibilities, model_data, galaxy_image_dict, galaxy_model_visibilities_dict, model_visibilities_of_galaxies_list

Already cached (no change needed): inversion (AG+AL), galaxy_model_image_dict (AL), model_images_of_planes_list (AL)

Key Files

• PyAutoArray/autoarray/fit/fit_dataset.py — base class, 12 changes
• PyAutoGalaxy/autogalaxy/imaging/fit_imaging.py — AG FitImaging, 8 changes
• PyAutoGalaxy/autogalaxy/interferometer/fit_interferometer.py — 6 changes
• PyAutoLens/autolens/imaging/fit_imaging.py — AL FitImaging, 5 changes

Testing

• pytest test_autoarray/ + pytest test_autogalaxy/ + pytest test_autolens/
• autolens_profiling/quick_update/imaging.py and imaging_delaunay.py
• Smoke tests across workspaces

Original Prompt

Click to expand starting prompt

Audit uncached @Property on Fit classes — cache immutable computed results

FitImaging.model_data is a plain @property that recomputes the entire inversion pipeline on every access. Any code that touches fit.model_data more than once (residuals, chi-squared, visualization, aggregator scripts) pays the full cost again. For a Delaunay source with 1500 mesh pixels, each recomputation is ~5-20s.

This was discovered during quick-update rendering profiling: accessing fit.model_data once and fit.subtracted_images_of_planes_list once (which internally accesses model_data again) cost 13.7s just from redundant recomputation.

FitImaging is constructed once and never mutated — the tracer, dataset, and settings are fixed at construction time. All computed properties should be safe to cache.

🤖 Generated with Claude Code

[Jammy2211]

Shipped — 2026-05-27

PRs

• perf: cache expensive @property on Fit classes #341
• perf: cache expensive @property on Fit classes PyAutoGalaxy#462
• perf: cache expensive @property on Fit classes PyAutoLens#548

Summary

Changed 38 expensive @property to @functools.cached_property across FitDataset, FitImaging, and FitInterferometer. Properties like model_data, residual_map, chi_squared_map, and log_likelihood no longer recompute on every access. For Delaunay inversions this eliminates ~13s of redundant recomputation per visualization pass. 2072 tests pass, smoke tests confirm no regressions (8 pre-existing failures unrelated to this change).