In MNE time frequency is improtant for any further analysis for MEG EEG and fMRI
here I address some from the toturials just for a remark.
1. Time Frequency functions in MNE
all the output data will be the following two format you could get the details by X.data: AverageTFR(info, data, times, freqs, nave[, …]) Container for Time-Frequency data. EpochsTFR(info, data, times, freqs[, …]) Container for Time-Frequency data on epochs. %% here most we will use psd_welch for psd and tfr_morlet and ftr_multitaper for time frequency analysis. csd_epochs(epochs[, mode, fmin, fmax, fsum, …]) Estimate cross-spectral density from epochs. psd_welch(inst[, fmin, fmax, tmin, tmax, …]) Compute the power spectral density (PSD) using Welch’s method. psd_multitaper(inst[, fmin, fmax, tmin, …]) Compute the power spectral density (PSD) using multitapers. fit_iir_model_raw(raw[, order, picks, tmin, …]) Fit an AR model to raw data and creates the corresponding IIR filter. tfr_morlet(inst, freqs, n_cycles[, use_fft, …]) Compute Time-Frequency Representation (TFR) using Morlet wavelets. tfr_multitaper(inst, freqs, n_cycles[, …]) Compute Time-Frequency Representation (TFR) using DPSS tapers. tfr_stockwell(inst[, fmin, fmax, n_fft, …]) Time-Frequency Representation (TFR) using Stockwell Transform. %% the ourput of the functions could be 'pow' 'itc''complex' 'csd'
epochall['left'].plot_psd(fmin=2., fmax=40.) %% 2-40 hz
power.plot(, baseline=(-0.5, 0), mode='logratio', title=power.ch_names) power.plot(, baseline=(-0.5, 0), mode='logratio', title=power.ch_names) power.plot_topo(baseline=(-0.5, 0), mode='logratio', title='Average power') fig, axis = plt.subplots(1, 2, figsize=(7, 4)) power.plot_topomap(ch_type='eeg', tmin=0.5, tmax=1.5, fmin=8, fmax=12, baseline=(0,0.5), mode='logratio', axes=axis, title='Alpha', vmax=0.45, show=False) power.plot_topomap(ch_type='eeg', tmin=0.5, tmax=1.5, fmin=13, fmax=25, baseline=(0, 0.5), mode='logratio', axes=axis, title='Beta', vmax=0.45, show=False) mne.viz.tight_layout() plt.show()
itc.plot_topo(title='Inter-Trial coherence', vmin=0., vmax=1., cmap='Reds')
when there isa parameter: average ==True
we will get the time frequency data shape as(22, 24, 417) (channel fres timepoint)
if they are FALSe the time freq datashape will be (9, 22, 24, 417) here 9 means epochs
then we have a general idear how to put the time frequency data in to cnn and rnn for deep learning part.
for (9, 22, 24, 417)
here we have 9 samples
each sample with 22channels
each channels has 24 freq unit with 417 time length.
in keras the LSTM model has the input as [batch_szie, timestep,input_dim], how could we map the time freq into LSTM cell ?
we have one more dim than the standerd input data.
anyway we have lot of choices such like a stacked RNN just as the brain singel is going on there.
I had tried some RNN in ERP data, sorry that the result is terrible far less than LDA SVM with CSP.
anyway keep diging.