def test_shape_output():
sel_funcs = ['mean', 'variance', 'kurtosis', 'pow_freq_bands']
features = extract_features(data, sfreq, sel_funcs, n_jobs=1)
features_as_df = extract_features(data, sfreq, sel_funcs,
n_jobs=1, return_as_df=True)
expected_shape = (n_epochs, (3 + 5) * n_channels)
assert_equal(features.shape, expected_shape)
assert_equal(features, features_as_df.values)
def test_feature_names_pow_freq_bands():
_data = data[:, :2, :] # keep only 2 channels for the sake of simplicity
selected_funcs = ['pow_freq_bands']
fb1 = np.array([[4., 8.], [30., 70.]])
fb2 = {'theta': [4, 8], 'low-gamma': np.array([30, 70])}
_fb = [fb1, fb2]
ratios_col_names1 = [
'ch0_band0/band1', 'ch0_band1/band0', 'ch1_band0/band1',
'ch1_band1/band0'
]
ratios_col_names2 = [
'ch0_theta/low-gamma', 'ch0_low-gamma/theta', 'ch1_theta/low-gamma',
'ch1_low-gamma/theta'
]
_ratios_names = [ratios_col_names1, ratios_col_names2]
pow_col_names1 = ['ch0_band0', 'ch0_band1', 'ch1_band0', 'ch1_band1']
pow_col_names2 = [
'ch0_theta', 'ch0_low-gamma', 'ch1_theta', 'ch1_low-gamma'
]
_pow_names = [pow_col_names1, pow_col_names2]
for fb, ratios_names, pow_names in zip(_fb, _ratios_names, _pow_names):
# With `ratios = 'only'`:
df_only = extract_features(_data,
sfreq,
selected_funcs,
funcs_params={
'pow_freq_bands__ratios': 'only',
'pow_freq_bands__freq_bands': fb
},
return_as_df=True)
assert_equal(df_only.columns.get_level_values(1).values, ratios_names)
# With `ratios = 'all'`:
df_all = extract_features(_data,
sfreq,
selected_funcs,
funcs_params={
'pow_freq_bands__ratios': 'all',
'pow_freq_bands__freq_bands': fb
},
return_as_df=True)
assert_equal(
df_all.columns.get_level_values(1).values,
pow_names + ratios_names)
# With `ratios = None`:
df = extract_features(_data,
sfreq,
selected_funcs,
funcs_params={
'pow_freq_bands__ratios': None,
'pow_freq_bands__freq_bands': fb
},
return_as_df=True)
assert_equal(df.columns.get_level_values(1).values, pow_names)
def test_optional_params():
features1 = extract_features(data, sfreq, ['spect_edge_freq'],
{'spect_edge_freq__edge': [0.6]})
features2 = extract_features(data, sfreq, ['spect_edge_freq'],
{'spect_edge_freq__edge': [0.5, 0.95]})
features3 = extract_features(data, sfreq, ['svd_fisher_info'],
{'svd_fisher_info__tau': 5})
assert_equal(features1.shape[-1], n_channels)
assert_equal(features3.shape[-1], n_channels)
assert_equal(features2.shape[-1], features1.shape[-1] * 2)
def test_user_defined_feature_function():
# User-defined feature function
@nb.jit()
def top_feature(arr, gamma=3.14):
return np.sum(np.power(gamma * arr, 3) - np.power(arr / gamma, 2),
axis=-1)
# Valid feature extraction
selected_funcs = ['mean', ('top_feature', top_feature)]
feat = extract_features(data, sfreq, selected_funcs)
assert_equal(feat.shape, (n_epochs, 2 * n_channels))
# Changing optional parameter ``gamma`` of ``top_feature``
feat2 = extract_features(data,
sfreq,
selected_funcs,
funcs_params={'top_feature__gamma': 1.41})
assert_equal(feat2.shape, (n_epochs, 2 * n_channels))
# Invalid feature extractions
with assert_raises(ValueError):
# Alias is already used
extract_features(data, sfreq, ['variance', ('mean', top_feature)])
# Tuple is not of length 2
extract_features(
data, sfreq,
['variance', ('top_feature', top_feature, data[:, ::2])])
# Invalid type
extract_features(data, sfreq, ['mean', top_feature])
def test_channel_naming():
ch_names = ['CHANNEL%s' % i for i in range(n_channels)]
ch_names[:4] = ['Cz', 'FCz', 'P1', 'CP1']
selected_funcs = ['app_entropy']
df = extract_features(
data, sfreq, selected_funcs, ch_names=ch_names, return_as_df=True)
expected_col_names = [('app_entropy', ch_name) for ch_name in ch_names]
assert df.columns.values.tolist() == expected_col_names
ch_names.append('CHANNEL%s' % n_channels)
with assert_raises(ValueError):
# incorrect number of channel names
df = extract_features(
data, sfreq, selected_funcs, ch_names=ch_names, return_as_df=True)
def online_pipe(self, data: NDArray) -> NDArray:
"""
The method get the data as ndarray with dimensions of (n_channels, n_samples).
The method returns the features for the given data.
:param data: ndarray with the shape (n_channels, n_samples)
:return: ndarray with the shape of (1, n_features)
"""
# Prepare the data to MNE functions
data = data.astype(np.float64)
# Filter the data (band-pass only)
data = mne.filter.filter_data(data, l_freq=8, h_freq=30, sfreq=self.eeg.sfreq, verbose=False)
# Laplacian
data = self.eeg.laplacian(data, self.eeg.get_board_names())
# Normalize
scaler = StandardScaler()
data = scaler.fit_transform(data.T).T
# Extract features
funcs_params = {'pow_freq_bands__freq_bands': np.array([8, 10, 12.5, 30])}
selected_funcs = ['pow_freq_bands', 'variance']
X = extract_features(data[np.newaxis], self.eeg.sfreq, selected_funcs, funcs_params)[0]
return X
def generate_mne_features_of_one_file(signals, sfreq, selected_funcs,
func_params, epoch_duration_s,
max_abs_val, agg_mode):
if agg_mode in ["none", "None", None]:
agg_mode = None
else:
getattr(np, agg_mode)
epochs = split_into_epochs(signals=signals,
sfreq=sfreq,
epoch_duration_s=epoch_duration_s)
mask = reject_windows_with_outliers(epochs, outlier_value=max_abs_val)
epochs = epochs[mask == False]
if epochs.size == 0:
logging.warning("removed all epochs due to outliers")
return None, None
# generate features implemented in mne_features
features = extract_features(epochs,
sfreq,
selected_funcs,
funcs_params=func_params,
return_as_df=True)
# aggregate over dimension of epochs
if agg_mode:
features = agg_mode(features, axis=0)
return features
def test_feature_extractor():
selected_funcs = ['app_entropy']
extractor = FeatureExtractor(sfreq=sfreq, selected_funcs=selected_funcs)
expected_features = extract_features(data, sfreq, selected_funcs)
assert_almost_equal(expected_features, extractor.fit_transform(data))
with assert_raises(ValueError):
FeatureExtractor(
sfreq=sfreq, selected_funcs=selected_funcs,
params={'app_entropy__metric': 'sqeuclidean'}).fit_transform(data)
def test_wrong_params():
with assert_raises(ValueError):
# Negative sfreq
extract_features(data, -0.1, ['mean'])
with assert_raises(ValueError):
# Unknown alias of feature function
extract_features(data, sfreq, ['power_freq_bands'])
with assert_raises(ValueError):
# No alias given
extract_features(data, sfreq, list())
with assert_raises(ValueError):
# Passing optional arguments with unknown alias
extract_features(data, sfreq, ['higuchi_fd'], {'higuch_fd__kmax': 3})
def test_feature_names_spect_slope():
n_chans = 2 # keep only 2 channels for the sake of simplicity
_data = data[:, :n_chans, :]
selected_funcs = ['spect_slope']
stats = ['intercept', 'slope', 'MSE', 'R2']
col_names = [
'ch%s_%s' % (ch, stat) for ch in range(n_chans) for stat in stats
]
df = extract_features(_data, sfreq, selected_funcs, return_as_df=True)
assert_equal(df.columns.get_level_values(1).values, col_names)
def extract_features(eeg: EEG, trials: List[RawArray],
features: List[str]) -> np.ndarray:
# Convert RawArray to ndarray
trials_ndarray = list(map(lambda x: x.get_data(), trials))
# Convert to 3d matrix
trials_ndarray = to_3d_matrix(trials_ndarray)
# Return features
return feature_extraction.extract_features(trials_ndarray,
sfreq=eeg.sfreq,
selected_funcs=features)
def test_channel_naming_bivariate(selected_func, include_diag):
ch_names = ['CHANNEL%s' % i for i in range(n_channels)]
ch_names[:4] = ['Cz', 'FCz', 'P1', 'CP1']
func_params = {selected_func + '__include_diag': include_diag}
df = extract_features(
data, sfreq, [selected_func], func_params, ch_names=ch_names,
return_as_df=True)
expected_col_names = [
(selected_func, ch_names[i] + '-' + ch_names[j])
for s, i, j in _idxiter(n_channels, include_diag=include_diag)]
assert df.columns.values.tolist() == expected_col_names
def extract_features_data(data,
selected_features,
sfreq=256,
funcs_params=None):
data_T = np.transpose(data, axes=[0, 2, 1])
#data_T = np.nan_to_num(data_T)
extracted_features = extract_features(X=data_T,
selected_funcs={selected_features},
funcs_params=funcs_params,
sfreq=sfreq,
return_as_df=True,
n_jobs=-1)
return extracted_features
def test_feature_names_quantile():
n_chans = 2 # keep only 2 channels for the sake of simplicity
_data = data[:, :n_chans, :]
selected_funcs = ['quantile']
q = [0.25, 0.75]
col_names = [
'ch%s_%s' % (ch, i) for ch in range(n_chans) for i in range(len(q))
]
df = extract_features(_data,
sfreq,
selected_funcs,
funcs_params={'quantile__q': q},
return_as_df=True)
assert_equal(df.columns.get_level_values(1).values, col_names)
def test_generic_features_names():
n_chans = 2 # keep only 2 channels for the sake of simplicity
_data = data[:, :n_chans, :]
selected_funcs = ([
'mean', 'variance', 'std', 'ptp_amp', 'skewness', 'kurtosis', 'rms',
'quantile', 'hurst_exp', 'app_entropy', 'samp_entropy', 'decorr_time',
'hjorth_mobility_spect', 'hjorth_complexity_spect', 'hjorth_mobility',
'hjorth_complexity', 'higuchi_fd', 'katz_fd', 'zero_crossings',
'line_length', 'spect_entropy', 'svd_entropy', 'svd_fisher_info'
])
col_names = [(func, 'ch%s' % ch) for func in selected_funcs
for ch in range(n_chans)]
df = extract_features(_data, sfreq, selected_funcs, return_as_df=True)
assert df.columns.to_list() == col_names
def test_channel_naming_pow_freq_bands():
ch_names = ['CHANNEL%s' % i for i in range(n_channels)]
ch_names[:4] = ['Cz', 'FCz', 'P1', 'CP1']
selected_funcs = ['pow_freq_bands']
func_params = {
'pow_freq_bands__freq_bands': np.array([[0, 2], [10, 20]]),
'pow_freq_bands__ratios': 'only'
}
df = extract_features(
data, sfreq, selected_funcs, func_params, ch_names=ch_names,
return_as_df=True)
expected_col_names = [
('pow_freq_bands', f'{ch_name}_band{i}/band{j}')
for ch_name in ch_names for _, i, j in _idxiter(2, triu=False)]
assert df.columns.values.tolist() == expected_col_names
def get_features(self, channels: List[str], selected_funcs: List[str],
notch: float = 50, low_pass: float = 4, high_pass: float = 48) -> NDArray:
"""
Returns features of all data since last call to get_board_data method.
:return features: NDArray of shape (1, n_features)
"""
# Get the raw data
data = self.get_raw_data(ch_names=channels)
# Filter
data = self.filter_data(data, notch, low_pass, high_pass)
# Extract features
features = extract_features(data.get_data()[0][np.newaxis], self.sfreq, selected_funcs)
return features
def test_feature_names_energy_freq_bands():
_data = data[:, :2, :] # keep only 2 channels for the sake of simplicity
selected_funcs = ['energy_freq_bands']
fb1 = np.array([[4., 8.], [30., 70.]])
fb2 = {'theta': [4, 8], 'low-gamma': np.array([30, 70])}
_fb = [fb1, fb2]
expected_names1 = ['ch0_band0', 'ch0_band1', 'ch1_band0', 'ch1_band1']
expected_names2 = ['ch0_theta', 'ch0_low-gamma',
'ch1_theta', 'ch1_low-gamma']
_expected_names = [expected_names1, expected_names2]
for fb, feat_names in zip(_fb, _expected_names):
df = extract_features(
_data, sfreq, selected_funcs,
funcs_params={'energy_freq_bands__freq_bands': fb},
return_as_df=True)
assert_equal(df.columns.get_level_values(1).values, feat_names)
def preprocess_dataset(output_dir):
ch_names = np.array([
'A1', 'A2', 'C3', 'C4', 'CZ', 'F3', 'F4', 'F7', 'F8', 'FP1', 'FP2',
'FZ', 'O1', 'O2', 'P3', 'P4', 'PZ', 'T3', 'T4', 'T5', 'T6'
])
data_paths = glob.glob('/storage/inria/viovene/tuh_data/**/*.edf',
recursive=True)
np.random.shuffle(data_paths)
train_paths, test_paths = train_test_split(data_paths)
data = {}
for dataset_name, data_paths in [('train', train_paths),
('test', test_paths)]:
sfreqs = []
xs = []
ys = []
for path in data_paths:
f = mne.io.read_raw_edf(path)
cleaned_ch_names = np.array([
c.replace('EEG ', '').replace('-REF', '') for c in f.ch_names
])
ch_idxs = np.array(
[np.where(cleaned_ch_names == ch)[0][0] for ch in ch_names])
sfreq = f.info['sfreq']
if sfreq != 250.0:
continue
sfreqs.append(sfreq)
x = f.get_data()
x = x[ch_idxs, :]
rnd_start_idx = np.random.randint(
int(2 * 60 * sfreq), int(x.shape[1] - (2 * 60 * sfreq)))
x = x[:, rnd_start_idx:int(rnd_start_idx + 60 * sfreq)]
xs.append(x[np.newaxis, :, :])
label = 'abnormal' in path
ys.append(label)
x = np.concatenate(xs, axis=0)
y = np.array(ys)
selected_funcs = {'mean', 'ptp_amp', 'std'}
x = extract_features(x, sfreqs[0], selected_funcs)
data[dataset_name] = {'x': x, 'y': y}
for dataset_name in data:
for k in ['x', 'y']:
path = os.path.join(output_dir, f'{k}_{dataset_name}.npy')
np.save(path, data[dataset_name][k])
def test_feature_names_spect_edge_freq():
n_chans = 2 # keep only 2 channels for the sake of simplicity
_data = data[:, :n_chans, :]
selected_funcs = ['spect_edge_freq']
_edges = [None, [.5], [.5, .9]]
for edge in _edges:
if edge is None:
edge = [.5]
col_names = [
'ch%s_%s' % (ch, i) for ch in range(n_chans)
for i in range(len(edge))
]
df = extract_features(_data,
sfreq,
selected_funcs,
funcs_params={'spect_edge_freq__edge': edge},
return_as_df=True)
assert_equal(df.columns.get_level_values(1).values, col_names)
def test_feature_names_wavelet_coef_energy(wavelet_name='db4'):
n_chans = 2 # keep only 2 channels for the sake of simplicity
_data = data[:, :n_chans, :]
selected_funcs = ['wavelet_coef_energy']
# number of coefficients of the DWT
wavelet = pywt.Wavelet(wavelet_name)
levdec = min(pywt.dwt_max_level(_data.shape[-1], wavelet.dec_len), 6)
col_names = [
'ch%s_%s' % (ch, i) for ch in range(n_chans) for i in range(levdec)
]
df = extract_features(
_data,
sfreq,
selected_funcs,
funcs_params={'wavelet_coef_energy__wavelet_name': wavelet_name},
return_as_df=True)
assert_equal(df.columns.get_level_values(1).values, col_names)
def test_optional_params_func_with_numba():
sel_funcs = ['higuchi_fd']
features1 = extract_features(data, sfreq, sel_funcs,
{'higuchi_fd__kmax': 5})
n_features1 = features1.shape[-1]
assert_equal(n_features1, n_channels)
import numpy as np
from mne_features.feature_extraction import extract_features
from moabb.datasets import physionet_mi
if __name__ == '__main__':
# get dataset
ds = physionet_mi.PhysionetMI()
raw = ds.get_data([2])[2]['session_0']['run_4'].pick_channels(['C3', 'C4'])
events = mne.events_from_annotations(raw)
s_freq = raw.info['sfreq']
# get x and save to file
X = mne.Epochs(raw, events[0]).get_data()
# extract features
extract_features()
params = {
'pow_freq_bands__freq_bands': np.arange(1, int(s_freq / 2), 1),
}
selected_funcs = {'mean', 'ptp_amp', 'std', 'pow_freq_bands'}
features_array = mne_features.feature_extraction.extract_features(
X, s_freq, selected_funcs, params)
# save features to file
np.savetxt('../data/mne/1/features.csv', features_array, delimiter=',')
# get y and save to file
y = np.asarray([e[2] for e in events[0]])
np.savetxt('../data/mne/1/stimulus_vectors.csv', y)
def test_njobs():
sel_funcs = ['app_entropy']
features = extract_features(data, sfreq, sel_funcs, n_jobs=-1)
expected_shape = (n_epochs, n_channels)
assert_equal(features.shape, expected_shape)
# Read epochs
epochs = mne.Epochs(raw, events, event_id, tmin, tmax, picks=picks, proj=True,
baseline=None, preload=True)
labels = epochs.events[:, -1]
# get MEG and EEG data
data = epochs.get_data()
###############################################################################
# Prepare for the classification task:
pipe = Pipeline([('scaler', StandardScaler()),
('lr', LogisticRegression(random_state=42, solver='lbfgs'))])
y = labels
###############################################################################
# Classification using features (mean, peak-to-peak amplitude,
# standard deviation). See :ref:`api_documentation` for full list of supported
# features.
selected_funcs = {'mean', 'ptp_amp', 'std'}
X_new = extract_features(data, raw.info['sfreq'], selected_funcs)
kf = KFold(n_splits=3, random_state=42)
scores = cross_val_score(pipe, X_new, y, scoring='accuracy', cv=kf)
###############################################################################
# Print the cross-validation score:
print('Cross-validation accuracy score = %1.3f (+/- %1.5f)' % (np.mean(scores),
np.std(scores)))
