[gbeane]

Summary

Window features are now the largest share of feature compute time (~59% in a recent jabs-init --timing-report run). Profiling one identity showed the std_dev, skew, and kurtosis window operations together taking ~28% of window time, and each independently recomputes the same per-window count, mean, centered values, and variance for every column.

• window_stats.window_std_skew_kurtosis(values, window) computes all three from one set of shared terms (_central_moment_terms). np_skew and np_kurtosis now use the same helpers, so the arithmetic is shared rather than duplicated. The standard deviation is sqrt of the same biased variance np.nanstd computes.
• Feature._window_standard computes the three through it in one pass (_shared_moment_features), keeping the original operation order. An operation is only routed through the shared path if it still uses its default implementation, so a subclass that overrides one of them keeps its own.
• The public window_std_dev / window_skew / window_kurtosis functions are unchanged.

Performance

tests/data/sample_pose_est_v6.h5, get_window_features(10) for all 4 identities, best of 3:

	time
before	2.388 s
after	2.023 s (-15%)

Feature values

Unchanged — no FEATURE_VERSION bump. Verified by dumping every window feature (54,936 arrays) for every identity in the v5 and v6 sample pose files at window sizes 5 and 10 on main and on this branch, and comparing for exact equality including key order.

Testing

• window_std_skew_kurtosis is bit-identical to the three separate ops on the existing test series (scattered nan, nan runs wider than the window, constant, all-nan, mostly zero) at three window sizes.
• Feature.window results equal each window op run on its own, in op order; a subclass overriding skew keeps its implementation.
• uv run pytest --ignore=tests/ui/test_viewer_window.py: 2090 passed, 7 skipped. tests/ui/test_viewer_window.py aborts the run with SIGTRAP after test_video_without_pose_file_opens_without_overlays locally on main too, so it's unrelated.

[Copilot]

🟡 Changes recommended

The shared reduction changes standard-deviation results for windows containing infinities from NaN to zero or infinity, altering feature values and requiring a fix with infinity-case equivalence tests.

1 open issue

• Preserve NaN results for infinite windows in window standard deviation · New

What changed in this PR

This PR reduces window-feature computation time by sharing intermediate calculations for standard deviation, skew, and kurtosis in JABS feature extraction.

Changes:

• Adds a combined window-statistics function and reusable central-moment helpers.
• Routes default moment operations through the shared calculation while preserving subclass overrides and output order.
• Adds equivalence tests for combined statistics and feature-level integration.

File	Description
tests/​feature_extraction/​window_operations/​test_window_stats.py	Tests combined statistics against separate operations.
tests/​feature_extraction/​test_feature_base_class.py	Tests output equivalence, operation order, and subclass overrides.
src/​jabs/​feature_extraction/​window_operations/​window_stats.py	Adds shared moment calculations and the combined statistics function.
src/​jabs/​feature_extraction/​feature_base_class.py	Integrates shared calculations into standard window features.

_{Review effort: Balanced}

[gbeane]

Good catch — fixed in 02c2b76. The std dev now sums the squared deviations like np.nanstd: only values that were nan in the input are zeroed, so a nan from inf - inf still propagates (skew/kurtosis keep their original nansum behavior, which was already shared). Added all-inf, mixed ±inf, and inf+nan series to the equivalence test at three window sizes (the previous commit fails all 9 cases); the 54,936-array comparison against main on the sample pose files is still exact.