CSP.plot_filters() is actually plotting the transpose of filters

[hongjiang-ye]

Describe the bug

The plot_filters() method of mne.decoding.CSP is not plotting filters, but actually plotting the transpose of filters. Using version 0.21.2.

Steps to reproduce

Use the code from Motor imagery decoding from EEG data using the Common Spatial Pattern (CSP) and change the last line of the second code block csp.plot_patterns to csp.plot_filters, and the result is:

And in mne.decoding.CSP, the each row of filters_ attribute is one CSP filter, see: https://github.com/mne-tools/mne-python/blob/maint/0.21/mne/decoding/csp.py#L184. But if we set the second row of filters_ to be all zeros, the figure almost has no change:

csp.filters_[1, :] = 0  # the coeffs in the second filter(row) are all 0 
csp.plot_filters(epochs.info, ch_type='eeg', units='Patterns (AU)', size=1.5)

Instead, if we set the second column of filters_ to be all zeros, then the "second filter" CSP1 becomes empty:

csp.filters_[:, 1] = 0  # the second column of filters_
csp.plot_filters(epochs.info, ch_type='eeg', units='Patterns (AU)', size=1.5)

So plot_filters is actually plotting the transpose of filters (the columns of filters_), not the csp filters (the rows of filters_).

To fix it

I think just change this line of code https://github.com/mne-tools/mne-python/blob/maint/0.21/mne/decoding/csp.py#L479 to filters = EvokedArray(self.filters_.T, info, tmin=0) could fix this issue. I tried the following code and the figure looks more reasonable to be CSP filters:

csp.filters_ = csp.filters_.T
csp.plot_filters(epochs.info, ch_type='eeg', units='Patterns (AU)', size=1.5)

BTW, the figure from plot_patterns method is correct.

[cbrnr]

LGTM. Can you submit a PR?

[hongjiang-ye]

Submitted: #8580. @cbrnr