Closed agramfort closed 10 years ago
possible data:
http://www.physionet.org/physiobank/database/eegmmidb/ see http://www.physionet.org/physiobank/database/eegmmidb/S001/
data is in edf format.
data from http://bbci.de/competition/iv/#download
cannot be downloaded without password....
You may be interested in this: https://github.com/SCoT-dev/scot-data but it is only a small data set in dull old .mat file format.
thanks
would you consider updating our CSP example (or create a new one) with a nice motor imagery dataset?
Also note that we plan to add custom 'from_data' constructors very soon that would allow you to create Raw / Epochs from e.g. mat files
I will look into it. This might be a nice way to get started with MNE :)
Uh.. I think I discovered an evil bug in RawEDF._read_segment
.
There is a stimulation channel containing ascii encoded strings, which is resampled before being incorrectly parsed. Interested in a fix?
sure pull request welcome
nice ID: #1234
@agramfort Finally, I've patched together a CSP example with motor imagery data: https://github.com/kazemakase/mne-python/blob/csp-bci-example/examples/decoding/plot_decoding_csp_eeg.py
I noticed that the CSP returns normalized band power (mean, std). Is there a reason for this particular normalization? We usually use log-bandpower for classification because this makes the features more normally distributed.
Another curiosity is that the CSP does not work with the empirical covariance. Eigenvalue decomposition complains that the covariance matrix is not positive definite. Using Ledoit-Wolf shrinkage fixes that problem. However, it is strange that the empirical covariance does not work here.
@agramfort https://github.com/agramfort Finally, I've patched together a CSP example with motor imagery data: https://github.com/kazemakase/mne-python/blob/csp-bci-example/examples/decoding/plot_decoding_csp_eeg.py
nice. Please open a PR asap.
I noticed that the CSP returns normalized band power (mean, std). Is the a reason for this particular normalization? We usually use log-bandpower for classification because this makes the features more normally distributed.
no reason just ignorance :) I don't work much with CSP. Thanks for the tip ;)
to PR #1267
now that we have support for EEG file formats it should be easy to add a BCI dataset with motor imagery that would be processed with CSP.
any volunteer?