Repository navigation
perf: cache immutable @property on Fit classes #340
Description
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
@propertymethods onFitDataset(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, andblurred_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 changesPyAutoGalaxy/autogalaxy/imaging/fit_imaging.py— AG FitImaging, 8 changesPyAutoGalaxy/autogalaxy/interferometer/fit_interferometer.py— 6 changesPyAutoLens/autolens/imaging/fit_imaging.py— AL FitImaging, 5 changes
Testing
pytest test_autoarray/+pytest test_autogalaxy/+pytest test_autolens/autolens_profiling/quick_update/imaging.pyandimaging_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
Activity
Jammy2211 commentedon May 27, 2026
Shipped — 2026-05-27
PRs
Summary
Changed 38 expensive
@propertyto@functools.cached_propertyacross FitDataset, FitImaging, and FitInterferometer. Properties likemodel_data,residual_map,chi_squared_map, andlog_likelihoodno 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).