Multichannel extraction¶
extract_features_mc handles multi-sensor recordings: one sample holds
C synchronous channels of a shared length.
Shapes¶
| Input | Output | Notes |
|---|---|---|
3D (n_samples, n_channels, length), C-contiguous float64 |
(n_samples, C * n_plan + n_cross) |
Zero-copy borrow like the 2D path |
List of 2D (C, length_i) float64 |
same, per-sample widths | Ragged lengths across samples; channel count must agree |
n_plan is the profile/feature count; n_cross is pairs * 4 + 4 when
cross=True and C >= 2 (else 0). Pairs run (0,1), (0,2), … capped by
max_pairs (default 8).
Column order and names¶
Per-channel blocks first, then cross columns — exactly
feature_names_mc(n_channels, ...):
ch{c}__{feature}for channelc, features in plan order.cross_corr_peak__ch{i}_ch{j},cross_corr_lag__ch{i}_ch{j},cross_cov__ch{i}_ch{j},cross_corr_coef__ch{i}_ch{j}per pair.cross__mean_abs_corr,cross__max_abs_corr,cross__eigen_spread,cross__coherence_meansummaries.
extract_features_mc_df returns the same matrix labeled; its columns are
asserted equal to feature_names_mc(...) in tests/test_multichannel.py.
import numpy as np
import kymora
X = np.random.default_rng(0).standard_normal((100, 4, 1000))
df = kymora.extract_features_mc_df(X, cross=True)
# Ragged: different recording lengths per sample
ragged = [np.random.default_rng(i).standard_normal((4, n)) for i, n in enumerate([500, 1000, 750])]
out = kymora.extract_features_mc(ragged)
Notes¶
- Per-channel values are bit-identical to running
extract_featureson that channel alone (tested). - Only float64 input is supported; float32 and non-raw views raise a clear
ValueError(same limitation as the batch f32 path). nan_policy="raise"reportssample {s} channel {c}; propagate follows the standard NaN contract per channel.- Non-contiguous input follows the global rule (
contiguous="error"by default,"copy"opt-in).