Environment
- meegkit 0.1.9, Python 3.11, pymanopt installed
Description
ASR with method='riemann' consistently applies reconstruction to the wrong channels — the lowest-variance (clean) ones instead of the highest-variance (noisy) ones. Providing clean calibration data to fit() has no effect on the output.
Steps to reproduce
import numpy as np
from meegkit.asr import ASR
rng = np.random.default_rng(42)
n_samples, sfreq = 50000, 256.0
clean = rng.normal(0, 5e-6, (4, n_samples))
clean[0] += rng.normal(0, 30e-6, n_samples) # ch 0 noisy
clean[3] += rng.normal(0, 30e-6, n_samples) # ch 3 noisy
cal = clean[:, :10000]
for method in ("euclid", "riemann"):
asr = ASR(sfreq=sfreq, cutoff=5.0, method=method)
asr.fit(cal)
rec = asr.transform(clean)
diff = np.abs(clean - rec)
thresh = np.median(diff, axis=1) + 3 * np.median(
np.abs(diff - np.median(diff, axis=1, keepdims=True)), axis=1)
print(f"{method}: {(thresh * 1e6).round(2).tolist()}")
Environment
Description
ASRwithmethod='riemann'consistently applies reconstruction to the wrong channels — the lowest-variance (clean) ones instead of the highest-variance (noisy) ones. Providing clean calibration data tofit()has no effect on the output.Steps to reproduce