Defer the eager scipy.sparse and scipy.spatial imports (~280 ms off import)

[Jammy2211]

Summary

import autoarray drops from 464.4 ms to 183.7 ms — a 281 ms saving, ~60% of the import cost. Medians of 15 runs each, Python 3.13, dev extras, same interpreter and machine.

This is the follow-up filed off the pynufft removal (#475), whose measurement turned up the real import-time cost.

The filed premise was wrong, in the same way the pynufft one was

The task named derivative_util.py:30's module-scope from scipy.sparse import csr_matrix as the target, on the evidence that scipy.sparse costs ~0.11 s of every import. Deferring it changed nothing — scipy.sparse was never being imported from there. Traced with a sys.meta_path hook:

autoarray/__init__.py:80
  -> inversion/mesh/mesh_geometry/delaunay.py:2   import scipy.spatial
    -> scipy/spatial/__init__.py:111              from ._kdtree import *
      -> scipy/spatial/_kdtree.py:4               from ._ckdtree import cKDTree

scipy.spatial (134 ms) pulls scipy.sparse (154 ms) in transitively. The csr_matrix import was riding on a subtree already paid for by something else — exactly the shape of the pynufft error, where a 0.19 s "cost" turned out to be 95% shared scipy.sparse. Deferring scipy.spatial too is what actually removes both, and is required to meet the task's own acceptance criterion.

That also means the win is ~280 ms rather than the ~100 ms the task predicted, because scipy.spatial's own cost comes off as well.

API Changes

None — internal changes only. No public symbol, signature, or behaviour changes; only the point at which two third-party modules are imported.

Test Plan

• python -X importtime -c "import autoarray" | grep scipy.sparse — empty
• python -X importtime -c "import autoarray" | grep scipy.spatial — empty
• import autoarray: 464.4 ms → 183.7 ms (medians of 15; main min/max 445.1/511.9, branch 172.8/204.9)
• pytest test_autoarray — 1179 passed
• Downstream against this branch: PyAutoGalaxy 1103 passed, 1 skipped; PyAutoLens 532 passed, 1 skipped

Changes by file

inversion/mesh/mesh_geometry/delaunay.py

Drops the module-scope import scipy.spatial. Two of its three use sites (lines 19 and 179 on main) already had local imports — this deferral had been started and left half-done; only voronoi_neighbors_from needed one added. The "scipy.spatial.Voronoi" return annotation is already a string literal, so it needs no import at definition time.

operators/derivative_util.py, operators/coarse_interp_util.py

Move from scipy.sparse import csr_matrix into the four functions that use it — derivative_1st_operators_from, derivative_2nd_operators_from, forward_difference_operators_from, coarse_interp_matrix_from. Not load-bearing for import time on its own any more, but it keeps scipy.sparse off the import path independently of what scipy.spatial happens to pull in, so this does not silently regress if scipy reshuffles its internals.

On the deferral pattern

Every use site is inside a function, so a plain local import suffices. No module-level cache as in transformer.py:_load_nufftax() — that one needed a cache only because unpickled instances in multiprocessing workers never re-run __init__. Function-local imports run on every call and hit sys.modules after the first, so the cost is a dict lookup.

---

Generated by Claude Code