I have some epoched data that I created in python (using mne.Epochs) and
then saved as a .fif file using the save() method of mne.Epochs. After
saving it, I did some other processing, and now I'd like to load them
again. But I'm actually not sure how to load these objects?
mne.fiff.read_evoked doesn't seem to do it (I get an error 'Could not find
evoked data'; the full traceback is below).
Also, is it possible to filter an mne.Epochs object? I didn't filter my raw
data because I wanted to do ICA on the epochs. But now, as far as I can
tell there is not a built-in filter() method for Epochs like there is for
Raw, and mne.filter.low_pass_filter() seems to be a low-level function so
I'm not sure if I should be calling it directly or not.
(My epochs are much larger than the time window I'm actually interested in,
so I think it should be ok to filter; but if that's still not recommended,
then another solution I could try is to run ICA on the epochs, apply those
ICA weights back onto the raw data, filter the raw data, and then epoch
again.)
this is not supported by the API because you should know what you do,
when you do it.
If you want to do it the way to do it is to access the epochs data (in
preload mode).
data = epochs.get_date()
n_epochs, n_channels, n_samples = data.shape
data = data.reshape(n_channels, n_epochs * n_samples)
# for example
mne.filter.low_pass_filter(data, 1, 45, copy=False)
epochs._data = data.reshape(n_epochs, n_channels, n_samples)
But this will only be valid if your data are sufficiently highpass
filtered (otherwise artifacts due to epochsing).
# ICA
is there a certain reason why you prefer to do ICA before filtering?
In my experience ICA will yield better results on filtered data
(removes e.g high-frequency noise and drifts).
Also you can speed up estimation time by not passing each sample using
the decim parameter having filtered the data. For separating signals
between 1 and 45 Hz you don't need 500 samples per second, which of
course depends on the total number of samples.
I often do the decim trick when working with raw data (it internally
decimates a copy of your data which is passed to ICA, not your
original data).
Regarding ICA and filtering, I just didn't do filtering because in the past
I haven't (I have always left off filtering until as late as possible, just
so I have the option of doing stats on unfiltered data if I want), but
actually I haven't systematically compared whether ICA works better on
filtered vs. unfiltered versions of my data, so it's definitely worth
looking into. As for why I did ICA on epoched data rather than raw, it's
just because I didn't want to pass in noisy data that's not from the actual
task (e.g., parts of the recording where the participant was taking a break
or talking to me, etc.), and the easiest way to do that was just to only
use the [large] epochs around my critical events. But I took epochs that
are each several seconds long, so that I can safely low-pass filter them
and then chop out the actual epoch of interest from the middle later.
As to ICA, It's perfectly valid to run it on epochs for the reasons
you mentioned.
My pitch was that low and highpass filtering should be done before,
unless you know better.
The reason is that not doing so you might produce spike like peaks in
your signals which maybe 'misinterpreted' as e.g. ECG peaks by ICA,
since ICA is not run on single trials but on the concatenated channels
by (samples * epochs) time series.