def get_data_2a(data_name, n_classes, num_channels=22):
# # # Reads in raw edf (or gdf) file from BCI Competition 2a and returns the signal, time array,
# # # and start of each (labeled) motor imagery event
# Returns: signal, time array, events
freq = 250
raw = read_raw_edf(data_name,
preload=True,
stim_channel='auto',
verbose='WARNING')
events = find_edf_events(raw)
events.pop(0)
time = events.pop(0)
events1 = events.pop(0)
events2 = events.pop(0)
events3 = events.pop(0)
# raw_train.plot_psd(area_mode='range', tmax=10.0)
raw.filter(7., 30., fir_design='firwin', skip_by_annotation='edge')
signal = np.transpose(raw.get_data()[:num_channels])
events = np.transpose(np.vstack([time, events1, events2, events3]))
time = raw.times * freq
return np.asarray(signal), time, events
def manage_edf_events(raw):
# find_edf_events returns the list of five elements:
# n_events - the number of all elements (this element
# excluded in order to transform list to array)
# pos - beginning of the events in samples
# typ - the event identifiers
# chn - the associated channels (0 for all)
# dur - duration of the events
event = find_edf_events(raw)
event1 = [event[1][:]]
event2 = [event[2][:]]
event3 = [event[3][:]]
events = np.vstack((event1, event3, event2))
events = events.transpose()
return events
def get_label_data_2a(data_name,
n_classes,
num_channels=22,
remove_rest=True,
training_data=True,
reuse_data=False,
mult_data=False,
noise_data=False):
freq = 250
raw = read_raw_edf(data_name,
preload=True,
stim_channel='auto',
verbose='WARNING')
events = find_edf_events(raw)
events.pop(0)
time = events.pop(0)
events1 = events.pop(0)
events2 = events.pop(0)
events3 = events.pop(0)
# raw_train.plot_psd(area_mode='range', tmax=10.0)
raw.filter(7., 30., fir_design='firwin', skip_by_annotation='edge')
signal = np.transpose(raw.get_data()[:num_channels])
events = np.transpose(np.vstack([time, events1, events2, events3]))
time = raw.times * freq
if training_data:
signal_out, labels = label_data_2a(signal,
time,
events,
remove_rest,
n_classes,
freq,
reuse_data=reuse_data,
mult_data=mult_data,
noise_data=noise_data)
# if remove_rest:
signal = signal_out
else:
labels = np.zeros(signal.shape[0])
return np.asarray(signal), labels
def load_data(file_path):
event_id = dict(left_hand=769, right_hand=770, both_feet=771, tongue=772)
tmin, tmax = -0.5, 4
raw = read_raw_edf(file_path, eog= [22, 23, 24], stim_channel="auto", preload=True)
raw.filter(0.5, 38.)
events = find_edf_events(raw)
events = np.stack((events[1], events[3], events[2]), 1).astype(np.int64)
picks = pick_types(raw.info, meg=False, eeg=True, stim=False, eog=False,
exclude='bads')
epochs_train = Epochs(raw, events, event_id, tmin, tmax, proj=True, picks=picks,
baseline=None, preload=True)
labels = epochs_train.events[:, -1] - 769
indices = np.arange(labels.shape[0])
np.random.shuffle(indices)
return epochs_train.get_data()[:, :, :, np.newaxis][indices], labels[indices]
def get_data_2a(data_name, n_classes, num_channels=22):
freq = 250
raw = read_raw_edf(data_name,
preload=True,
stim_channel='auto',
verbose='WARNING')
events = find_edf_events(raw)
events.pop(0)
time = events.pop(0)
events1 = events.pop(0)
events2 = events.pop(0)
events3 = events.pop(0)
# raw_train.plot_psd(area_mode='range', tmax=10.0)
raw.filter(7., 30., fir_design='firwin', skip_by_annotation='edge')
signal = np.transpose(raw.get_data()[:num_channels])
events = np.transpose(np.vstack([time, events1, events2, events3]))
time = raw.times * freq
return np.asarray(signal), time, events
def get_data_2a(data_name, n_classes, num_channels=22):
# # # Collects data from edf file and returns it as an np array along with the events
# # # This has been specifically designed to handle data from the BCI Competition IV 2a dataset
# Returns: signal as np array, time array, and the events/prompts
freq = 250
raw = read_raw_edf(data_name, preload=True, stim_channel='auto', verbose='WARNING')
events = find_edf_events(raw)
events.pop(0)
time = events.pop(0)
events1 = events.pop(0)
events2 = events.pop(0)
events3 = events.pop(0)
# raw_train.plot_psd(area_mode='range', tmax=10.0)
raw.filter(7., 30., fir_design='firwin', skip_by_annotation='edge')
signal = np.transpose(raw.get_data()[:num_channels])
events = np.transpose(np.vstack([time, events1, events2, events3]))
time = raw.times * freq
return np.asarray(signal), time, events
def prepare_eeg_csv():
root = './data'
classes = [
'img_lhand', 'img_rhand', 'img_bfeet', 'img_tongue', 'img_none'
]
class_mapping = {
'img_lhand': 769,
'img_rhand': 770,
'img_bfeet': 771,
'img_tongue': 772,
}
# load recording gdf and extract its headers
bci_record_names = []
for file_name in os.listdir(root):
bci_record_names.append(file_name)
# print('Filename: {}'.format(bci_record_names[0]))
# stim channel is required for gdf files
gdf_file = io.read_raw_edf(root + '/' + bci_record_names[0],
stim_channel=1,
preload=True)
recording_headers = io.find_edf_events(gdf_file)
recording_positions = recording_headers[1]
recording_types = recording_headers[2]
recording_durations = recording_headers[4]
# convert to dataframe (recording time series)
recording_ts = gdf_file.to_data_frame()
# drop EOG cols
eog_cols = ['EOG-left', 'EOG-central', 'EOG-right']
recording_ts.drop(eog_cols, axis=1, inplace=True)
# fill nans with 0
recording_ts.fillna(0.0, inplace=True)
# normalize
recording_ts = (recording_ts - recording_ts.min()) / recording_ts.std()
# extract class ranges
class_ranges = EEGPrep.get_class_ranges(class_mapping, recording_types,
recording_positions,
recording_durations)
# add a new column for labels
recording_ts_labeled = EEGPrep.label_recording(recording_ts,
class_ranges)
# fill all nans with 'none'
recording_ts_labeled['class_label'].fillna('img_none', inplace=True)
recording_ts_labeled['class_label'].replace(0.0,
'img_none',
inplace=True)
# convert to one-hot
recording_ts_oh = EEGDataUtils.one_hot(recording_ts_labeled,
['class_label'])
# shuffle data
recording_ts_oh = recording_ts_oh.sample(frac=1)
return recording_ts_oh