Baseline
June 14, 2021
1 Hackathon baseline
We provide here a simple pipeline to read the data, train a Tangent Space Classifier and try naive
transfer between sessions.
[1]: %matplotlib inline
import os
import mne
import pandas as pd
from mne.externals.pymatreader import read_mat
import numpy as np
import matplotlib.pyplot as plt
import itertools
from glob import glob
import pyriemann
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import cross_val_score
from sklearn.pipeline import make_pipeline
from sklearn.metrics import accuracy_score
Here set the data_path to corresponding path on your computer
[8]: data_path = '/home/dcas/l.darmet/data/contest/comeptition_done'
n_subs = 4
n_sessions = 2
diff = ['MATBeasy', 'MATBmed', 'MATBdiff']
Read channels names and position
[10]: electrodes = pd.read_csv(data_path + '/Electrodes/
,→chan_locs_standard',header=None, sep ='\t', names=['ch_names','x','y','z'])
electrodes.head()
[10]: ch_names x y z
0 Fp1 -29.4370 83.917 -6.990
1
1 Fz 0.3122 58.512 66.462
2 F3 -50.2440 53.111 42.192
3 F7 -70.2630 42.474 -11.420
4 FT9 -84.0760 14.567 -50.429
Covariance estimation For robust covariance estimation, we take advantage of shrinkage. Here
the Oracle Approximating Shrinkage (OAS) is used. #### Classifier We use a simple Logistic
Regression (with a non-optimized L2 penalty) on Tangent Space Features, extracted with Pyrie-
mann toolbox. #### Channel selection A manual and naive EEG channel selection is performed
to use 13 electrodes, mostly frontal.
[ ]: lr = LogisticRegression(C=1/10.)
clf = make_pipeline(pyriemann.estimation.Covariances(estimator='oas'),
pyriemann.classification.TSclassifier(clf=lr))
ch_slice = ['F7', 'F5', 'F3', 'F1', 'F2', 'F4', 'F6', 'AF3', 'AFz', 'AF4',␣
,→'FP1', 'FP2', 'FPz']
1.1 Single subject epochs classification
[ ]: for sub_n, session_n in itertools.product(range(n_subs), range(n_sessions)):
epochs_data = []
labels = []
for lab_idx, level in enumerate(diff):
sub = 'P{0:02d}'.format(sub_n+1)
sess = f'S{session_n+1}'
path = os.path.join(os.path.join(data_path, sub), sess) + f'/eeg/
,→alldata_sbj{str(sub_n+1).zfill(2)}_sess{session_n+1}_{level}.set'
# Read the epoched data with MNE
epochs = mne.io.read_epochs_eeglab(path, verbose=False)
# You could add some pre-processing here with MNE
# We will just select some channels (mostly frontal ones)
epochs = epochs.drop_channels(list(set(epochs.ch_names) -␣
,→set(ch_slice)))
# Get the data and concatenante with others MATB levels
tmp = epochs.get_data()
epochs_data.extend(tmp)
labels.extend([lab_idx]*len(tmp))
epochs_data = np.array(epochs_data)
labels = np.array(labels)
# Compute classification accuracy with 5-folds cross validation
acc = cross_val_score(clf, X=epochs_data, y=labels, cv=5)
2
print(f'Subject {sub} and session {session_n+1}: mean accuracy of {round(np.
,→ mean(acc), 2)} with a standard deviation of {round(np.std(acc), 2)}')
1.2 Transfer from session 1 to session 2 for P01
For subject P01, a model is trained on session 1 and directly used for epochs of session 2
[ ]: sub_n = 0
[ ]: session_n = 0
epochs_data = []
labels = []
for lab_idx, level in enumerate(diff):
sub = 'P{0:02d}'.format(sub_n+1)
sess = f'S{session_n+1}'
path = os.path.join(os.path.join(data_path, sub), sess) + f'/eeg/
,→alldata_sbj{str(sub_n+1).zfill(2)}_sess{session_n+1}_{level}.set'
# Read the epoched data with MNE
epochs = mne.io.read_epochs_eeglab(path, verbose=False)
# You could add some pre-processing here with MNE
# We will just select some channels (mostly frontal ones)
epochs = epochs.drop_channels(list(set(epochs.ch_names) - set(ch_slice)))
# Get the data and concatenante with others MATB levels
tmp = epochs.get_data()
epochs_data.extend(tmp)
labels.extend([lab_idx]*len(tmp))
epochs_data = np.array(epochs_data)
labels = np.array(labels)
# Train the model on all epochs from session 1
clf.fit(epochs_data, labels)
[ ]: session_n = 1
epochs_data = []
labels = []
for lab_idx, level in enumerate(diff):
sub = 'P{0:02d}'.format(sub_n+1)
sess = f'S{session_n+1}'
path = os.path.join(os.path.join(data_path, sub), sess) + f'/eeg/
,→alldata_sbj{str(sub_n+1).zfill(2)}_sess{session_n+1}_{level}.set'
# Read the epoched data with MNE
tmp = mne.io.read_epochs_eeglab(path, verbose=False)
# You could add some pre-processing here with MNE
# We will just select some channels (mostly frontal ones)
3
epochs = epochs.drop_channels(list(set(epochs.ch_names) - set(ch_slice)))
# Get the data and concatenante with others MATB levels
tmp = epochs.get_data()
epochs_data.extend(tmp)
labels.extend([lab_idx]*len(tmp))
epochs_data = np.array(epochs_data)
labels = np.array(labels)
# Use trained model to predict for all epochs of session 2 and compute accuracy
y_pred = clf.predict(epochs_data)
acc = accuracy_score(labels, y_pred)
print(f'Subject {sub} and transfer from session 1 to 2: mean accuracy of␣
,→{round(acc, 2)}.')
Generate a CSV to submit
[ ]: submission = pd.DataFrame({'epochID':np.arange(len(y_pred)), 'prediction' :␣
,→y_pred})
submission.head()
[ ]: submission.to_csv("submission.csv",header=True,index=False)
