Julius/semesterproject_lecture_eeg
1
0
Fork 0
Code Issues Pull Requests Releases Activity

Initial commit

Browse Source
This commit is contained in:
Julius Voggesberger
2021-03-27 14:58:49 +01:00
commit fc4eec6ac7
20 changed files with 1613 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

239
decoding_tf_analysis.py Normal file
View File

@@ -0,0 +1,239 @@
import math
import mne
import numpy as np
import matplotlib.pyplot as plt
from mne.decoding import SlidingEstimator, cross_val_multiscore, Vectorizer, Scaler
from mne.time_frequency import tfr_morlet
from sklearn.linear_model import LogisticRegression
from sklearn.pipeline import make_pipeline
from utils.file_utils import load_preprocessed_data, get_epochs
from utils.plot_utils import plot_tf_cluster, plot_oscillation_bands
VERBOSE_LEVEL = 'CRITICAL'
def events_to_labels(evts, events_dict, mask=None): # TODO Test schreiben
"""
Converts the event labels of epochs to class labels for classification
:param evts: the event labels to be converted
:param events_dict: a dictionary of event keys
:param mask: an optional label mask with 4-entries, where:
1. entry: 'face intact', 2. entry: 'car intact', 3. entry: 'face scrambled', 4. entry: 'face scrambled'
If None the entries are [0,1,0,1] i.e. all faces are in class 0 and all cars are in class 1
:return: The list of class labels
"""
events = evts.copy()
if mask is None:
mask = [0, 1, 0, 1]
for i in range(len(events)):
key = list(events_dict.keys())[list(events_dict.values()).index(events[i])]
k = int(key.split(':')[1])
if k < 41:
events[i] = mask[0] # Face intact
elif 40 < k < 81:
events[i] = mask[1] # Car intact
elif 100 < k < 141:
events[i] = mask[2] # Face scrambled
elif 140 < k < 181:
events[i] = mask[3] # Car scrambled
return events
def permutation_test(baseline, score, n_iter):
"""
An implementation of a permutation test for classification scores.
:param baseline: The classification scores of the baseline, i.e. selection by chance
:param score: The classification scores which are tested for significance
:param n_iter: number of permutations
:return: p-value
"""
all_data = np.concatenate((baseline, score))
# Base statistic. The statistic used here is the difference of means
given_diff = np.mean(score) - np.mean(baseline)
all_diffs = [given_diff]
# Permutation iterations
for i in range(n_iter):
# Create a permutation of indices and then use indices from index 0 to len(baseline) to get data for baseline.
# Analogously for scores
perm_indices = np.random.permutation(list(range(len(all_data))))
mean_diff = np.mean(all_data[perm_indices[len(baseline):]]) - np.mean(all_data[perm_indices[:len(baseline)]])
all_diffs.append(mean_diff)
p_val = len(np.where(np.asarray(all_diffs) >= given_diff)[0]) / (n_iter + 1)
return p_val
def decoding(dataset, filename, compute_metric=True, mask=None):
"""
Runs decoding over time for all subjects
:param dataset: The dataset for which the decoding is done
:param filename: filename of either the file from which the classifier scores will be loaded
or to which they will be saved
:param compute_metric: If True the classifier will be run, else the result will be loaded from a precomputed file
:param mask: an optional label mask with 4-entries, where:
1. entry: 'face intact', 2. entry: 'car intact', 3. entry: 'face scrambled', 4. entry: 'face scrambled'
If None the entries are [0,1,0,1] i.e. all faces are in class 0 and all cars are in class 1
"""
if mask is None:
mask = [0, 1, 0, 1]
times = None
time_scale = 1100
metric = []
p_values = []
if compute_metric:
# Computes classifier scores for all subjects
for i in range(1, 41):
subj = "0" + str(i)
if len(str(i)) == 1:
subj = "0" + subj
# Load data
raw = load_preprocessed_data(subj, dataset)
epochs, events_dict = get_epochs(raw, picks=mne.pick_types(raw.info, eeg=True, eog=False))
data = epochs.get_data()
labels = events_to_labels(epochs.events[:, 2], events_dict, mask)
# Classify
clf = make_pipeline(Scaler(epochs.info), Vectorizer(), LogisticRegression(solver='lbfgs'))
time_decode = SlidingEstimator(clf)
scores = cross_val_multiscore(time_decode, data, labels, cv=10, n_jobs=4)
metric.append(np.mean(scores, axis=0))
if times is None:
times = epochs.times
np.save('cached_data/decoding_data/' + filename, metric)
else:
# Dummy time which is created according to epoch.times
times = np.linspace(-0.09960938, 1, 1127)
metric = np.load('cached_data/decoding_data/' + filename + '.npy')
# Compute index of time point 0
index = math.floor((len(metric[0]) / time_scale) * 100)
baseline = np.array(metric[:index]).flatten()
plt.plot(np.linspace(-200, 1000, 1127), np.mean(metric, axis=0))
plt.ylabel('Accuracy (%)')
plt.xlabel('Time (ms)')
plt.title('Mean Accuracy over Subjects for Faces vs. Cars')
plt.show()
# Compute the permutation tests
for t in range(len(metric[0][index:])):
score_t = np.asarray(metric[:, t + index])
p = permutation_test(baseline, score_t, 100)
p_values.append(p)
if t % 50 == 0:
print(str(t) + " Out of " + str(len(metric[0][index:])))
plt.plot(times[index:], p_values)
plt.ylabel('P-Value')
plt.xlabel('Time (ms)')
plt.title('P-Values for Faces vs. Cars')
plt.show()
def create_tfr(raw, condition, freqs, n_cycles, response='induced', baseline=None):
"""
Compute the time frequency representation (TFR) of data for a given condition via morlet wavelets
:param raw: the data
:param condition: the condition for which to compute the TFR. Given as a list of tuples of the form (stimulus, condition) # TODO ambiguous use of condition
:param freqs: the frequencies for which to compute the TFR
:param n_cycles: the number of cycles used by the morlet wavelets
:param response: type of expected TFR. Can be total, induced or evoked. Default is induced
:param baseline: baseline used to correct the power. A tuple of the form (start, end).
Default is None and no baseline correction will be applid
:return: The TFR or the given data for a given condition. Has type AverageTFR
"""
epochs, _ = get_epochs(raw, condition, tmin=-0.2, tmax=1)
print(' ' + str(condition))
if response == 'total':
print(' Power Total')
power = tfr_morlet(epochs, freqs=freqs, n_cycles=n_cycles, return_itc=False, n_jobs=4)
elif response == 'induced':
print(' Power Induced')
power = tfr_morlet(epochs.subtract_evoked(), freqs=freqs, n_cycles=n_cycles, return_itc=False, n_jobs=4)
else:
print(' Power Evoked')
power_total = tfr_morlet(epochs, freqs=freqs, n_cycles=n_cycles, return_itc=False, n_jobs=4)
power_induced = tfr_morlet(epochs.subtract_evoked(), freqs=freqs, n_cycles=n_cycles, return_itc=False, n_jobs=4)
power = mne.combine_evoked([power_total, power_induced], weights=[1, -1])
# power.plot(picks='P7')
power.apply_baseline(mode='ratio', baseline=baseline)
# plot_oscillation_bands(power)
# power.plot(picks='P7')
return power
def time_frequency(dataset, filename, compute_tfr=True):
"""
Runs time frequency analysis
:param dataset: The dataset for which the decoding is done
:param filename: Filename of either the file from which the TFRs will be loaded
or to which they will be saved
:param compute_tfr: If True the TFRs will be created, else the TFRs will be loaded from a precomputed file
"""
# Parameters
# Frequency space (from, to, steps) -> Control frequency resolution : Between num=50-80 good for 1-50Hz
freqs = np.logspace(*np.log10([0.5, 50]), num=50) #
# Number of cycles -> Controls time resolution ? At ~freqs/2 good for high frequency resolution
n_cycles = freqs / 2 # 1 for high time resolution & freq smoothing, freqs/2 for high freq resolution & time smooth
# Baseline -> Should not go post-stimulus, i.e. > 0 -> Best ist pre-stimulus (e.g. -400 to -200ms)
baseline = [-0.5, 0]
cond1 = []
cond2 = []
times = None
if compute_tfr:
for i in range(1, 41):
subj = "0" + str(i)
if len(str(i)) == 1:
subj = "0" + subj
print("########## SUBJECT " + subj + " ##########")
# Load data
raw = load_preprocessed_data(subj, dataset)
raw.set_channel_types({'HEOG_left': 'eog', 'HEOG_right': 'eog', 'VEOG_lower': 'eog'})
raw.set_montage('standard_1020', match_case=False)
# Create the two conditions we want to compare
power_cond1 = create_tfr(raw, [('face', 'intact')], freqs, n_cycles, 'induced', (-0.2, 0))
print(' CONDITION 1 LOADED')
cond1.append(power_cond1)
power_cond2 = create_tfr(raw, [('face', 'scrambled'), ('car', None)], freqs, n_cycles, 'induced',
(-0.2, 0))
print(' CONDITION 2 LOADED')
cond2.append(power_cond2)
print(' DONE')
np.save('cached_data/tf_data/' + filename + '_cond1', cond1)
np.save('cached_data/tf_data/' + filename + '_cond2', cond2)
else:
cond1 = np.load('cached_data/tf_data/' + filename + '_cond1.npy', allow_pickle=True).tolist()
cond2 = np.load('cached_data/tf_data/' + filename + '_cond2.npy', allow_pickle=True).tolist()
if times is None:
times = cond1[0].times
mne.grand_average(cond2).plot(picks=['P7'], vmin=-3, vmax=3, title='Grand Average P7')
plot_oscillation_bands(mne.grand_average(cond1))
plot_oscillation_bands(mne.grand_average(cond2))
F, clusters, cluster_p_values, h0 = mne.stats.permutation_cluster_test(
[mne.grand_average(cond1).data, mne.grand_average(cond2).data], n_jobs=4, verbose='INFO',
seed=123)
plot_tf_cluster(F, clusters, cluster_p_values, freqs, times)
if __name__ == '__main__':
mne.set_log_level(verbose=VERBOSE_LEVEL)
ds = 'N170'
# decoding(ds, 'faces_vs_cars_100iters', False)
time_frequency(ds, 'face_intact_vs_all_0.1_50hz_ncf2', False)