nk.signal_recompose() is broken

[yanbingyu07]

I am trying testing the signal decompose/recompose functions with the example from the home page, as shown below:

import neurokit2 as nk
import numpy as np

# Create complex signal
signal = nk.signal_simulate(duration=10, frequency=1)  # High freq
signal += 3 * nk.signal_simulate(duration=10, frequency=3)  # Higher freq
signal += 3 * np.linspace(0, 2, len(signal))  # Add baseline and linear trend
signal += 2 * nk.signal_simulate(duration=10, frequency=0.1, noise=0)  # Non-linear trend
signal += np.random.normal(0, 0.02, len(signal))  # Add noise
nk.signal_plot(signal)  # Visualize signal

# Decompose signal using Empirical Mode Decomposition (EMD)
components = nk.signal_decompose(signal, method='emd')
nk.signal_plot(components)  # Visualize components

# Recompose merging correlated components
recomposed = nk.signal_recompose(components, threshold=0.99)
nk.signal_plot(recomposed)  # Visualize components

But the signal_recompose() function was broken. The traceback shows line#103 of signal_recompose.py with error:

File c:\Users\scssh059\AppData\Local\Programs\Python\Python313\Lib\site-packages\neurokit2\signal\signal_recompose.py:103, in _signal_recompose_wcorr(components, threshold, metric)
    100 wcorr = _signal_recompose_get_wcorr(components, show=False)
    102 # Find clusters in correlation matrix
--> [103](file:///C:/Users/scssh059/AppData/Local/Programs/Python/Python313/Lib/site-packages/neurokit2/signal/signal_recompose.py:103) pairwise_distances = scipy.cluster.hierarchy.distance.pdist(wcorr, metric=metric)
    104 linkage = scipy.cluster.hierarchy.linkage(pairwise_distances, method="complete")
    105 threshold = threshold * pairwise_distances.max()

AttributeError: module 'scipy.cluster.hierarchy' has no attribute 'distance'

The version of scipy is 1.18.0 on my pc, which is running Windows11.

[DerAndereJohannes]

Thank you for the report! It looks like scipy 1.18.0 did some cleanup of their code here scipy/scipy#24060. That being said, it looks like we were piggybacking off of the imports from scipy.cluster.hierarchy and we should have been going off of the scipy.spatial.distance directly. I have made this change and already put it into dev.

With this said, you now have the choice of staying on the current stable neurokit (0.2.13) and restricting your scipy version to < 1.18.0, or if you like you can install the current dev branch of neurokit and use the newer version of scipy (pip install -U git+https://github.com/neuropsychology/NeuroKit@dev if using pip or I think its something like uv add "neurokit2 @ git+https://github.com/neuropsychology/NeuroKit.git@dev" for uv projects).

Let me know if you come into any additional issues!

Thanks!

[yanbingyu07]

Johannes,
Thanks for your reply. I did some research about scipy after my last post, and found out the same as you described above. I did change the line#103 of signal_recompose.py to use scipy.spatial.distance.pdist(), which worked out ok.

But I have another question. The recompose function seems returning a set of component waveforms, instead of one waveform composed from the modified components. Can the compose function return one combined waveform?

Thank you for the great tool.