Runtime import

[BalzaniEdoardo]

Summary

• Move heavy imports (scipy.signal, scipy.spatial, xarray) inside functions to reduce import pynapple time
• Import time reduced from ~920ms to ~390ms (~2.4x faster)

Background

As developers, we discussed using lazy_loader.load from scientific-python/lazy_loader but decided against it for this package because:

1. Early adoption risk - The tool is relatively new and may not be mature enough for production use
2. Limited gain vs. breaking change risk - The potential benefits don't outweigh the risk of introducing breaking changes to the import system

We recognize that the current pattern of moving imports inside functions is suboptimal (slight overhead on first function call, repeated import statements). However, we defer adoption of lazy_loader until it proves itself stable across the ecosystem. We already have it as a dependency in nemos, which will serve as a testing ground.

Changes

• scipy.signal imports moved inside functions in:
  • core/time_series.py (smooth, decimate methods)
  • core/_core_functions.py (_convolve)
  • process/filtering.py (already done)
  • process/spectrum.py (compute_mean_power_spectral_density)
• scipy.spatial.distance.cdist import moved inside process/decoding.py
• xarray imports moved inside wrapper functions in:
  • process/decoding.py
  • process/tuning_curves.py (already done)
• Added test test_lazy_import_heavy_modules to ensure these modules aren't loaded at import time

Performance

Branch	Import time (avg)
dev	~920 ms
runtime-import	~390 ms

Speedup: ~2.4x faster (57% reduction)

Additional fix

• Fixed test_compute_1d_mutual_info and test_compute_2d_mutual_info tests in tests/test_tuning_curves.py
• Changed np.testing.assert_approx_equal to np.testing.assert_allclose (lines 1591, 1663)
• assert_approx_equal is for scalars only; numpy 2.4.2 (used in CI) enforces this strictly, causing TypeError: only 0-dimensional arrays can be converted to Python scalars

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
pynapple/core/_core_functions.py	85.18% <100.00%> (ø)
pynapple/core/time_series.py	93.60% <100.00%> (+<0.01%)	⬆️
pynapple/process/decoding.py	92.43% <100.00%> (+0.12%)	⬆️
pynapple/process/filtering.py	91.83% <100.00%> (+0.17%)	⬆️
pynapple/process/spectrum.py	87.03% <100.00%> (ø)
pynapple/process/tuning_curves.py	97.35% <100.00%> (+0.04%)	⬆️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[h-mayorquin]

For a similar reason, in SpikeInterface, we keep an action on github measuring import time for the different modules:

https://github.com/SpikeInterface/spikeinterface/blob/main/.github/workflows/test_imports.yml

https://github.com/SpikeInterface/spikeinterface/blob/main/.github/scripts/import_test.py

This avoids regression.

With time, I have come to love the simplicity of delayed imports with just an import x on the spot where it is needed. Honestly, the downsides are 99 % aesthetical but you learn to get over them.