Spectrogram calculation error when a state doesn't activate (`analysis/spectral.py`)

[scho97]

When calculating a multitaper spectrogram, the following error can arise:

RuntimeWarning: invalid value encountered in divide  psd_ /= fo_

which occurred in spectral.py (L1727-L1728). This affects subsequent analysis and plotting.

Here, the resulting psd_ value is an array of np.nan, and this turns out to be because fo_ = 0 and psd_ is an array of zeroes. This was a very specific case, occurring for a single subject within a model single run (so far). In such case, it seems like one state did not activate over entire time points.

=> print(psd_ / fo_)
[[nan nan nan ... nan nan nan]
 [nan nan nan ... nan nan nan]
 [nan nan nan ... nan nan nan]
 ...
 [nan nan nan ... nan nan nan]
 [nan nan nan ... nan nan nan]
 [nan nan nan ... nan nan nan]]
=> print(fo_)
0.0
=> print(psd_)
[[0. 0. 0. ... 0. 0. 0.]
 [0. 0. 0. ... 0. 0. 0.]
 [0. 0. 0. ... 0. 0. 0.]
 ...
 [0. 0. 0. ... 0. 0. 0.]
 [0. 0. 0. ... 0. 0. 0.]
 [0. 0. 0. ... 0. 0. 0.]]

For this, I wasn't sure if we should consider this as a poor model fit (and therefore train the model again) or we should fix it with something like

psd_ = np.nan_to_num(psd_)

as we did for the coherence in spectral.py (L855-L856).

[cgohil8]

I left this in because having no state activation probably indicates the HMM fit didn't go so well. Getting an error makes this ver obvious, otherwise the user might think everything's fine.

I guess would eventually realise the HMM fit's off based on the spectra being full of zeros.

Maybe we should add nan_to_num for the PSDs. @RukuangHuang what do you think?

[RukuangHuang]

nan_to_num will be good. But I think we should add a warning that warns the user of potential poor HMM fit. And maybe suggest solutions like training multiple times and get the lowert FE run.

[scho97]

Okay, I can make a pull request with nan_to_num and a detailed warning message, and perhaps both of you can review it?