ECE 401 Fall 2021 MP4: Ideal and Practical Filters

In this file, we're going to lowpass, bandpass, and highpass filter some EEG signals. The signals we'll be using are samples provided with the MNE package. It is recommended, but not required, that you first install MME. You can do so using the instructions on the MNE homepage. On a MAC you will go to a command window, and type

conda create --name=mne --channel=conda-forge mne
conda install --name=base nb_conda_kernels

If you don't want to install MNE, you can do the whole MP without it, by simply avoiding the blocks of code marked Requires MNE below.

import os
import numpy as np
import matplotlib.figure
import matplotlib.pyplot
%matplotlib inline
import importlib
import submitted

Data: With MNE

Here's how the data works. If you want to use MNE to visualize the data, you first need to install the mne environment as described above. Then, every time you want to run the MNE code on this page, you'll need to do the following steps:

1. In an operating system command window, type conda activate mne, and then jupyter-notebook. That should open up your Jupyter browser, just as if you had opened it from the GUI, except it will make the following steps possible (they wouldn't be possible if you used the GUI).
2. Open this file (mp4_notebook.ipynb).
3. Go to the top of this screen, and find the Kernel menu. Under that menu, choose Change Kernel -> Python [conda env:mne].

Now you should be able to run the following block of code without errors:

## REQUIRES MNE
import mne,os
os.makedirs(os.path.join(os.path.expanduser('~'),'mne_data'), exist_ok=True)
sample_data_folder = mne.datasets.sample.data_path()
sample_data_raw_file = os.path.join(sample_data_folder,'MEG','sample','sample_audvis_raw.fif')
raw = mne.io.read_raw_fif(sample_data_raw_file, preload=True, verbose=False)
data, times = raw.get_data(picks='eeg',return_times=True)
channel_indices_by_type = mne.io.pick.channel_indices_by_type(raw.info)
len(channel_indices_by_type['eeg'])
info = mne.pick_info(raw.info, channel_indices_by_type['eeg'])
print('According to the MNE Info, the Sampling Freq is %g'%(info['sfreq']))

Downloading archive MNE-sample-data-processed.tar.gz to /Users/jhasegaw/mne_data
Downloading https://files.osf.io/v1/resources/rxvq7/providers/osfstorage/59c0e26f9ad5a1025c4ab159?version=5&action=download&direct (1.54 GB)

Verifying hash 12b75d1cb7df9dfb4ad73ed82f61094f.
Decompressing the archive: /Users/jhasegaw/mne_data/MNE-sample-data-processed.tar.gz
(please be patient, this can take some time)
Successfully extracted to: ['/Users/jhasegaw/mne_data/MNE-sample-data']
According to the MNE Info, the Sampling Freq is 600.615

Data: With or Without MNE

If you don't want to use MNE, you can run the following block of code to load the repackaged data. You should run the following block whether or not you're using MNE, because some of the blocks that follow will use this version of the data.

import h5py
with h5py.File('data.hdf5','r') as f:
    onesecond = f['onesecond'][:]
    sfreq = f['sfreq'][0]
solutions = h5py.File('solutions.hdf5','r')
print('According to data.hdf5, sfreq is %g, nchannels is %d, nsamps is %d'%
     (sfreq, onesecond.shape[0], onesecond.shape[1]))

According to data.hdf5, sfreq is 600.615, nchannels is 60, nsamps is 600

1. Browsing the Data

EEG data consists of 60 different voltage signals, recorded in parallel from different electrodes on the scalp. Here is a function that you can use to choose three of the signals, and plot both the signals and their power spectra:

def plot_waves_and_spectra(signal, sfreq, nrows=3):
    nchannels, nsamps = signal.shape
    t = np.linspace(0,nsamps/sfreq,nsamps,endpoint=False)
    f = np.linspace(0,sfreq,nsamps,endpoint=False)
    fig = matplotlib.figure.Figure(figsize=(14,6))
    axs = fig.subplots(nrows,2)
    channels = np.arange(0,nchannels,int(nchannels/nrows))
    for plotrow in range(nrows):
        channel = channels[plotrow]
        axs[plotrow,0].plot(t,signal[channel,:])
        axs[plotrow,0].set_title('Signal, Channel %d'%(channel))
        axs[plotrow,0].set_ylabel('Volts')
        axs[plotrow,0].set_xlabel('Seconds')
        axs[plotrow,1].plot(f,20*np.log10(1e-3+np.abs(np.fft.fft(signal[channel,:]))))
        axs[plotrow,1].set_title('Spectrum, Channel %d'%(channel))
        axs[plotrow,1].set_xlabel('Hertz')
        axs[plotrow,1].set_ylabel('dB')
    return(fig)

plot_waves_and_spectra(onesecond,sfreq)

Browsing the Data: With MNE

If you have installed MNE, you can use the plot_topomap function to see how the voltages are distributed over the scalp at each time step. Here's a function that plots the voltages as a function of location, for 25 different time steps:

def plot_topomaps(data, info):
    nchannels, nsamps = data.shape
    fig = matplotlib.figure.Figure(figsize=(10,8))
    axs = fig.subplots(5,5)
    for row in range(5):
        for col in range(5):
            samp = int(nsamps*(row*5+col)/25)
            t = samp/info['sfreq']
            mne.viz.plot_topomap(data[:,samp],info,axes=axs[row,col],show=False)
            axs[row,col].set_title('%dms'%(int(np.round(1000*t))))
    return(fig)

Running it requires MNE. It requires that you have already run the MNE sample-loading code, up above, to load the info object that contains the spatial locations of the electrodes.

# REQUIRES MNE
plot_topomaps(onesecond,info)

<Figure size 432x288 with 0 Axes>

2. Running the MP

Now you're ready to run the MP. Each block of code, below, will first show you the distributed solution (distributed to you in the file solutions.hdf5), and will then run the code in your own file submitted.py, to test whether or not your code works.

def plot_digital_filter(h, sfreq):
    fig = matplotlib.figure.Figure(figsize=(14,6))
    axs = fig.subplots(3,1)
    axs[0].plot(h)
    axs[0].set_title('Impulse response $h[n]$')
    H = np.abs(np.fft.fft(h))
    omega = np.linspace(0,2*np.pi,len(h))
    axs[1].plot(omega,H)
    axs[1].set_title('Magnitude response $|H(\omega)|$')
    f = omega *  sfreq / (2*np.pi)
    axs[2].loglog(f[1:],1e-3+H[1:])
    axs[2].set_title('Log-log plot of magnitude vs. frequency in Hertz')
    return(fig)

2.1 todo_lpf_even

This function tests whether or not you can create a lowpass filter with a specified cutoff frequency, and with even length. The distributed solution contains a 200-sample lowpass filter with a cutoff of pi/4. Notice that the loglog plot doesn't quite know what to do with the aliased part of the filter, up near $\omega=2\pi$. It isn't really designed to handle the frequency responses of digital filters like this one.

plot_digital_filter(solutions['h_even'],sfreq)