def select_best_parcellation(parcellations,
clf,
avg_signals,
y,
n_jobs=1,
verbose=False):
"""
Returns the best parcellation of the parcellations given in arguments.
Parameters
----------
parcellations : (int list) list
list of all the parcellations possible for
the current iteration
clf : classifier
avg_signals : nparray
average_signals for each node
y : ndarray of shape = (n_samples)
n_jobs : int, optional
number of cpu to use
verbose : int, optional
does it really need explanations?
Returns
-------
parcellation : int list
best parcellation of the parcellations given in argument
"""
# Computing scores for each parcellation
scores = Parallel(n_jobs)(
delayed(cross_val.cross_val_score)(
estimator=clf,
X=avg_signals[:, j],
y=y,
#cv=cross_val.KFold(avg_signals.shape[0], 5),
n_jobs=1) for j in parcellations)
if verbose >= 2:
print " Scores of each parcellation for current iteration :"
print scores
print scores
scores = Parallel(n_jobs)(delayed(np.mean)(i) for i in scores)
indice = np.argmax(scores)
return parcellations[indice], scores[indice]
def select_best_parcellation(parcellations, clf, avg_signals, y, n_jobs=1,
verbose=False):
"""
Returns the best parcellation of the parcellations given in arguments.
Parameters
----------
parcellations : (int list) list
list of all the parcellations possible for
the current iteration
clf : classifier
avg_signals : nparray
average_signals for each node
y : ndarray of shape = (n_samples)
n_jobs : int, optional
number of cpu to use
verbose : int, optional
does it really need explanations?
Returns
-------
parcellation : int list
best parcellation of the parcellations given in argument
"""
# Computing scores for each parcellation
scores = Parallel(n_jobs)(delayed(cross_val.cross_val_score)
(estimator=clf, X=avg_signals[:, j], y=y,
#cv=cross_val.KFold(avg_signals.shape[0], 5),
n_jobs=1)
for j in parcellations)
if verbose >= 2:
print " Scores of each parcellation for current iteration :"
print scores
print scores
scores = Parallel(n_jobs)(delayed(np.mean)(i) for i in scores)
indice = np.argmax(scores)
return parcellations[indice], scores[indice]
def para_bmu_find(self, x, y, njb=1):
dlen = x.shape[0]
Y2 = None
Y2 = np.einsum('ij, ij->i', y, y)
bmu = None
b = None
# Here it finds BMUs for chunk of data in parallel
t_temp = time()
b = Parallel(n_jobs=njb, pre_dispatch='3 * n_jobs')(
delayed(chunk_based_bmu_find)
(x[i * dlen // njb:min((i + 1) * dlen // njb, dlen)], y, Y2)
for i in xrange(njb))
# print 'bmu finding: {} seconds '.format(round(time() - t_temp, 3))
t1 = time()
bmu = np.asarray(list(itertools.chain(*b))).T
# print 'bmu to array: {} seconds'.format(round(time() - t1, 3))
del b
return bmu
def parallel_func(func, n_jobs, verbose=5):
"""Return parallel instance with delayed function
Util function to use joblib only if available
Parameters
----------
func: callable
A function
n_jobs: int
Number of jobs to run in parallel
verbose: int
Verbosity level
Returns
-------
parallel: instance of joblib.Parallel or list
The parallel object
my_func: callable
func if not parallel or delayed(func)
n_jobs: int
Number of jobs >= 0
"""
try:
from scikits.learn.externals.joblib import Parallel, delayed
parallel = Parallel(n_jobs, verbose=verbose)
my_func = delayed(func)
if n_jobs == -1:
try:
import multiprocessing
n_jobs = multiprocessing.cpu_count()
except ImportError:
print "multiprocessing not installed. Cannot run in parallel."
n_jobs = 1
except ImportError:
print "joblib not installed. Cannot run in parallel."
n_jobs = 1
my_func = func
parallel = list
return parallel, my_func, n_jobs
def permutation_t_test(X, n_permutations=10000, tail=0, n_jobs=1):
"""One sample/paired sample permutation test based on a t-statistic.
This function can perform the test on one variable or
simultaneously on multiple variables. When applying the test to multiple
variables, the "tmax" method is used for adjusting the p-values of each
variable for multiple comparisons. Like Bonferroni correction, this method
adjusts p-values in a way that controls the family-wise error rate.
However, the permutation method will be more
powerful than Bonferroni correction when different variables in the test
are correlated.
Parameters
----------
X : array of shape [n_samples x n_tests]
Data of size number of samples (aka number of observations) times
number of tests (aka number of variables)
n_permutations : int or 'all'
Number of permutations. If n_permutations is 'all' all possible
permutations are tested (2**n_samples). It's the exact test, that
can be untractable when the number of samples is big (e.g. > 20).
If n_permutations >= 2**n_samples then the exact test is performed
tail : -1 or 0 or 1 (default = 0)
If tail is 1, the alternative hypothesis is that the
mean of the data is greater than 0 (upper tailed test). If tail is 0,
the alternative hypothesis is that the mean of the data is different
than 0 (two tailed test). If tail is -1, the alternative hypothesis
is that the mean of the data is less than 0 (lower tailed test).
n_jobs : int
Number of CPUs to use for computation.
Returns
-------
T_obs : array of shape [n_tests]
T-statistic observed for all variables
p_values : array of shape [n_tests]
P-values for all the tests (aka variables)
H0 : array of shape [n_permutations]
T-statistic obtained by permutations and t-max trick for multiple
comparison.
Notes
-----
A reference (among many) in field of neuroimaging:
Nichols, T. E. & Holmes, A. P. (2002). Nonparametric permutation tests
for functional neuroimaging: a primer with examples.
Human Brain Mapping, 15, 1-25.
Overview of standard nonparametric randomization and permutation
testing applied to neuroimaging data (e.g. fMRI)
DOI: http://dx.doi.org/10.1002/hbm.1058
"""
n_samples, n_tests = X.shape
do_exact = False
if n_permutations is 'all' or (n_permutations >= 2 ** n_samples - 1):
do_exact = True
n_permutations = 2 ** n_samples - 1
X2 = np.mean(X ** 2, axis=0) # precompute moments
mu0 = np.mean(X, axis=0)
dof_scaling = sqrt(n_samples / (n_samples - 1.0))
std0 = np.sqrt(X2 - mu0 ** 2) * dof_scaling # get std with var splitting
T_obs = np.mean(X, axis=0) / (std0 / sqrt(n_samples))
if do_exact:
perms = bin_perm_rep(n_samples, a=1, b=-1)[1:, :]
else:
perms = np.sign(0.5 - np.random.rand(n_permutations, n_samples))
try:
from scikits.learn.externals.joblib import Parallel, delayed
parallel = Parallel(n_jobs)
my_max_stat = delayed(_max_stat)
except ImportError:
print "joblib not installed. Cannot run in parallel."
n_jobs = 1
my_max_stat = _max_stat
parallel = list
if n_jobs == -1:
try:
import multiprocessing
n_jobs = multiprocessing.cpu_count()
except ImportError:
print "multiprocessing not installed. Cannot run in parallel."
n_jobs = 1
max_abs = np.concatenate(parallel(my_max_stat(X, X2, p, dof_scaling)
for p in np.array_split(perms, n_jobs)))
H0 = np.sort(max_abs)
scaling = float(n_permutations + 1)
if tail == 0:
p_values = 1.0 - np.searchsorted(H0, np.abs(T_obs)) / scaling
elif tail == 1:
p_values = 1.0 - np.searchsorted(H0, T_obs) / scaling
elif tail == -1:
p_values = 1.0 - np.searchsorted(H0, -T_obs) / scaling
return T_obs, p_values, H0
def source_induced_power(epochs, inverse_operator, bands, lambda2=1.0 / 9.0,
dSPM=True, n_cycles=5, df=1, use_fft=False,
baseline=None, baseline_mode='logratio', pca=True,
subtract_evoked=False, n_jobs=1):
"""Compute source space induced power
Parameters
----------
epochs: instance of Epochs
The epochs
inverse_operator: instance of inverse operator
The inverse operator
bands: dict
Example : bands = dict(alpha=[8, 9])
lambda2: float
The regularization parameter of the minimum norm
dSPM: bool
Do dSPM or not?
n_cycles: int
Number of cycles
df: float
delta frequency within bands
use_fft: bool
Do convolutions in time or frequency domain with FFT
baseline: None (default) or tuple of length 2
The time interval to apply baseline correction.
If None do not apply it. If baseline is (a, b)
the interval is between "a (s)" and "b (s)".
If a is None the beginning of the data is used
and if b is None then b is set to the end of the interval.
If baseline is equal ot (None, None) all the time
interval is used.
baseline_mode: None | 'logratio' | 'zscore'
Do baseline correction with ratio (power is divided by mean
power during baseline) or zscore (power is divided by standard
deviatio of power during baseline after substracting the mean,
power = [power - mean(power_baseline)] / std(power_baseline))
pca: bool
If True, the true dimension of data is estimated before running
the time frequency transforms. It reduces the computation times
e.g. with a dataset that was maxfiltered (true dim is 64)
subtract_evoked: bool
If True, the evoked component (average of all epochs) if subtracted
from each epochs.
n_jobs: int
Number of jobs to run in parallel
"""
if n_jobs == -1:
try:
import multiprocessing
n_jobs = multiprocessing.cpu_count()
except ImportError:
print "multiprocessing not installed. Cannot run in parallel."
n_jobs = 1
try:
from scikits.learn.externals.joblib import Parallel, delayed
parallel = Parallel(n_jobs)
my_compute_power = delayed(_compute_power)
except ImportError:
print "joblib not installed. Cannot run in parallel."
n_jobs = 1
my_compute_power = _compute_power
parallel = list
#
# Set up the inverse according to the parameters
#
epochs_data = epochs.get_data()
if subtract_evoked: # subtract with a copy not to touch epochs
epochs_data = epochs_data - np.mean(epochs_data, axis=0)
nave = len(epochs_data) # XXX : can do better when no preload
inv = prepare_inverse_operator(inverse_operator, nave, lambda2, dSPM)
#
# Pick the correct channels from the data
#
sel = [epochs.ch_names.index(name) for name in inv['noise_cov']['names']]
print 'Picked %d channels from the data' % len(sel)
print 'Computing inverse...',
#
# Simple matrix multiplication followed by combination of the
# three current components
#
# This does all the data transformations to compute the weights for the
# eigenleads
#
K = inv['reginv'][:, None] * reduce(np.dot,
[inv['eigen_fields']['data'],
inv['whitener'],
inv['proj']])
if pca:
U, s, Vh = linalg.svd(K)
rank = np.sum(s > 1e-8*s[0])
K = s[:rank] * U[:, :rank]
Vh = Vh[:rank]
print 'Reducing data rank to %d' % rank
else:
Vh = None
#
# Transformation into current distributions by weighting the
# eigenleads with the weights computed above
#
if inv['eigen_leads_weighted']:
#
# R^0.5 has been already factored in
#
# print '(eigenleads already weighted)...',
K = np.dot(inv['eigen_leads']['data'], K)
else:
#
# R^0.5 has to factored in
#
# print '(eigenleads need to be weighted)...',
K = np.sqrt(inv['source_cov']['data'])[:, None] * \
np.dot(inv['eigen_leads']['data'], K)
Fs = epochs.info['sfreq'] # sampling in Hz
stcs = dict()
src = inv['src']
for name, band in bands.iteritems():
print 'Computing power in band %s [%s, %s] Hz...' % (name, band[0],
band[1])
freqs = np.arange(band[0], band[1] + df / 2.0, df) # frequencies
Ws = morlet(Fs, freqs, n_cycles=n_cycles)
power = sum(parallel(my_compute_power(data, K, sel, Ws,
inv['source_ori'], use_fft, Vh)
for data in np.array_split(epochs_data, n_jobs)))
if dSPM:
# print '(dSPM)...',
power *= inv['noisenorm'][:, None] ** 2
# average power in band + mean over epochs
power /= len(epochs_data) * len(freqs)
# Run baseline correction
if baseline is not None:
print "Applying baseline correction ..."
times = epochs.times
bmin, bmax = baseline
if bmin is None:
imin = 0
else:
imin = int(np.where(times >= bmin)[0][0])
if bmax is None:
imax = len(times)
else:
imax = int(np.where(times <= bmax)[0][-1]) + 1
mean_baseline_power = np.mean(power[:, imin:imax], axis=1)
if baseline_mode is 'logratio':
power /= mean_baseline_power[:, None]
power = np.log(power)
elif baseline_mode is 'zscore':
power -= mean_baseline_power[:, None]
power /= np.std(power[:, imin:imax], axis=1)[:, None]
else:
print "No baseline correction applied..."
stc = SourceEstimate(None)
stc.data = power
stc.tmin = epochs.times[0]
stc.tstep = 1.0 / Fs
stc.lh_vertno = src[0]['vertno']
stc.rh_vertno = src[1]['vertno']
stc._init_times()
stcs[name] = stc
print '[done]'
return stcs
def induced_power(epochs, Fs, frequencies, use_fft=True, n_cycles=7, n_jobs=1):
"""Compute time induced power and inter-trial phase-locking factor
The time frequency decomposition is done with Morlet wavelets
Parameters
----------
epochs : array
3D array of shape [n_epochs, n_channels, n_times]
Fs : float
sampling Frequency
frequencies : array
Array of frequencies of interest
use_fft : bool
Compute transform with fft based convolutions or temporal
convolutions.
n_cycles : int
The number of cycles in the wavelet
n_jobs : int
The number of CPUs used in parallel. All CPUs are used in -1.
Requires joblib package.
Returns
-------
power : 2D array
Induced power (Channels x Frequencies x Timepoints).
Squared amplitude of time-frequency coefficients.
phase_lock : 2D array
Phase locking factor in [0, 1] (Channels x Frequencies x Timepoints)
"""
n_frequencies = len(frequencies)
n_epochs, n_channels, n_times = epochs.shape
# Precompute wavelets for given frequency range to save time
Ws = morlet(Fs, frequencies, n_cycles=n_cycles)
try:
import joblib
except ImportError:
print "joblib not installed. Cannot run in parallel."
n_jobs = 1
if n_jobs == 1:
psd = np.empty((n_channels, n_frequencies, n_times))
plf = np.empty((n_channels, n_frequencies, n_times), dtype=np.complex)
for c in range(n_channels):
X = np.squeeze(epochs[:, c, :])
psd[c], plf[c] = _time_frequency(X, Ws, use_fft)
else:
from joblib import Parallel, delayed
psd_plf = Parallel(n_jobs=n_jobs)(
delayed(_time_frequency)(np.squeeze(epochs[:, c, :]), Ws, use_fft)
for c in range(n_channels))
psd = np.zeros((n_channels, n_frequencies, n_times))
plf = np.zeros((n_channels, n_frequencies, n_times), dtype=np.complex)
for c, (psd_c, plf_c) in enumerate(psd_plf):
psd[c, :, :], plf[c, :, :] = psd_c, plf_c
psd /= n_epochs
plf = np.abs(plf) / n_epochs
return psd, plf
def single_trial_power(epochs,
Fs,
frequencies,
use_fft=True,
n_cycles=7,
baseline=None,
baseline_mode='ratio',
times=None,
n_jobs=1):
"""Compute time-frequency power on single epochs
Parameters
----------
epochs : instance Epochs | array of shape [n_epochs, n_channels, n_times]
The epochs
Fs : float
Sampling rate
frequencies : array-like
The frequencies
use_fft : bool
Use the FFT for convolutions or not.
n_cycles : float
The number of cycles in the Morlet wavelet
baseline: None (default) or tuple of length 2
The time interval to apply baseline correction.
If None do not apply it. If baseline is (a, b)
the interval is between "a (s)" and "b (s)".
If a is None the beginning of the data is used
and if b is None then b is set to the end of the interval.
If baseline is equal ot (None, None) all the time
interval is used.
baseline_mode : None | 'ratio' | 'zscore'
Do baseline correction with ratio (power is divided by mean
power during baseline) or zscore (power is divided by standard
deviatio of power during baseline after substracting the mean,
power = [power - mean(power_baseline)] / std(power_baseline))
times : array
Required to define baseline
n_jobs : int
The number of epochs to process at the same time
Returns
-------
power : list of 2D array
Each element of the list the the power estimate for an epoch.
"""
mode = 'same'
n_frequencies = len(frequencies)
n_epochs, n_channels, n_times = epochs.shape
# Precompute wavelets for given frequency range to save time
Ws = morlet(Fs, frequencies, n_cycles=n_cycles)
try:
from scikits.learn.externals.joblib import Parallel, delayed
parallel = Parallel(n_jobs)
my_cwt = delayed(cwt)
except ImportError:
print "joblib not installed. Cannot run in parallel."
n_jobs = 1
my_cwt = cwt
parallel = list
print "Computing time-frequency power on single epochs..."
power = np.empty((n_epochs, n_channels, n_frequencies, n_times),
dtype=np.float)
if n_jobs == 1:
for k, e in enumerate(epochs):
power[k] = np.abs(cwt(e, Ws, mode))**2
else:
# Precompute tf decompositions in parallel
tfrs = parallel(my_cwt(e, Ws, use_fft, mode) for e in epochs)
for k, tfr in enumerate(tfrs):
power[k] = np.abs(tfr)**2
# Run baseline correction
if baseline is not None:
if times is None:
raise ValueError('times parameter is required to define baseline')
print "Applying baseline correction ..."
bmin = baseline[0]
bmax = baseline[1]
if bmin is None:
imin = 0
else:
imin = int(np.where(times >= bmin)[0][0])
if bmax is None:
imax = len(times)
else:
imax = int(np.where(times <= bmax)[0][-1]) + 1
mean_baseline_power = np.mean(power[:, :, :, imin:imax], axis=3)
if baseline_mode is 'ratio':
power /= mean_baseline_power[:, :, :, None]
elif baseline_mode is 'zscore':
power -= mean_baseline_power[:, :, :, None]
power /= np.std(power[:, :, :, imin:imax], axis=3)[:, :, :, None]
else:
print "No baseline correction applied..."
return power
def source_induced_power(epochs,
inverse_operator,
bands,
lambda2=1.0 / 9.0,
dSPM=True,
n_cycles=5,
df=1,
use_fft=False,
baseline=None,
baseline_mode='logratio',
pca=True,
subtract_evoked=False,
n_jobs=1):
"""Compute source space induced power
Parameters
----------
epochs: instance of Epochs
The epochs
inverse_operator: instance of inverse operator
The inverse operator
bands: dict
Example : bands = dict(alpha=[8, 9])
lambda2: float
The regularization parameter of the minimum norm
dSPM: bool
Do dSPM or not?
n_cycles: int
Number of cycles
df: float
delta frequency within bands
use_fft: bool
Do convolutions in time or frequency domain with FFT
baseline: None (default) or tuple of length 2
The time interval to apply baseline correction.
If None do not apply it. If baseline is (a, b)
the interval is between "a (s)" and "b (s)".
If a is None the beginning of the data is used
and if b is None then b is set to the end of the interval.
If baseline is equal ot (None, None) all the time
interval is used.
baseline_mode: None | 'logratio' | 'zscore'
Do baseline correction with ratio (power is divided by mean
power during baseline) or zscore (power is divided by standard
deviatio of power during baseline after substracting the mean,
power = [power - mean(power_baseline)] / std(power_baseline))
pca: bool
If True, the true dimension of data is estimated before running
the time frequency transforms. It reduces the computation times
e.g. with a dataset that was maxfiltered (true dim is 64)
subtract_evoked: bool
If True, the evoked component (average of all epochs) if subtracted
from each epochs.
n_jobs: int
Number of jobs to run in parallel
"""
if n_jobs == -1:
try:
import multiprocessing
n_jobs = multiprocessing.cpu_count()
except ImportError:
print "multiprocessing not installed. Cannot run in parallel."
n_jobs = 1
try:
from scikits.learn.externals.joblib import Parallel, delayed
parallel = Parallel(n_jobs)
my_compute_power = delayed(_compute_power)
except ImportError:
print "joblib not installed. Cannot run in parallel."
n_jobs = 1
my_compute_power = _compute_power
parallel = list
#
# Set up the inverse according to the parameters
#
epochs_data = epochs.get_data()
if subtract_evoked: # subtract with a copy not to touch epochs
epochs_data = epochs_data - np.mean(epochs_data, axis=0)
nave = len(epochs_data) # XXX : can do better when no preload
inv = prepare_inverse_operator(inverse_operator, nave, lambda2, dSPM)
#
# Pick the correct channels from the data
#
sel = [epochs.ch_names.index(name) for name in inv['noise_cov']['names']]
print 'Picked %d channels from the data' % len(sel)
print 'Computing inverse...',
#
# Simple matrix multiplication followed by combination of the
# three current components
#
# This does all the data transformations to compute the weights for the
# eigenleads
#
K = inv['reginv'][:, None] * reduce(
np.dot, [inv['eigen_fields']['data'], inv['whitener'], inv['proj']])
if pca:
U, s, Vh = linalg.svd(K)
rank = np.sum(s > 1e-8 * s[0])
K = s[:rank] * U[:, :rank]
Vh = Vh[:rank]
print 'Reducing data rank to %d' % rank
else:
Vh = None
#
# Transformation into current distributions by weighting the
# eigenleads with the weights computed above
#
if inv['eigen_leads_weighted']:
#
# R^0.5 has been already factored in
#
# print '(eigenleads already weighted)...',
K = np.dot(inv['eigen_leads']['data'], K)
else:
#
# R^0.5 has to factored in
#
# print '(eigenleads need to be weighted)...',
K = np.sqrt(inv['source_cov']['data'])[:, None] * \
np.dot(inv['eigen_leads']['data'], K)
Fs = epochs.info['sfreq'] # sampling in Hz
stcs = dict()
src = inv['src']
for name, band in bands.iteritems():
print 'Computing power in band %s [%s, %s] Hz...' % (name, band[0],
band[1])
freqs = np.arange(band[0], band[1] + df / 2.0, df) # frequencies
Ws = morlet(Fs, freqs, n_cycles=n_cycles)
power = sum(
parallel(
my_compute_power(data, K, sel, Ws, inv['source_ori'], use_fft,
Vh)
for data in np.array_split(epochs_data, n_jobs)))
if dSPM:
# print '(dSPM)...',
power *= inv['noisenorm'][:, None]**2
# average power in band + mean over epochs
power /= len(epochs_data) * len(freqs)
# Run baseline correction
if baseline is not None:
print "Applying baseline correction ..."
times = epochs.times
bmin, bmax = baseline
if bmin is None:
imin = 0
else:
imin = int(np.where(times >= bmin)[0][0])
if bmax is None:
imax = len(times)
else:
imax = int(np.where(times <= bmax)[0][-1]) + 1
mean_baseline_power = np.mean(power[:, imin:imax], axis=1)
if baseline_mode is 'logratio':
power /= mean_baseline_power[:, None]
power = np.log(power)
elif baseline_mode is 'zscore':
power -= mean_baseline_power[:, None]
power /= np.std(power[:, imin:imax], axis=1)[:, None]
else:
print "No baseline correction applied..."
stc = SourceEstimate(None)
stc.data = power
stc.tmin = epochs.times[0]
stc.tstep = 1.0 / Fs
stc.lh_vertno = src[0]['vertno']
stc.rh_vertno = src[1]['vertno']
stc._init_times()
stcs[name] = stc
print '[done]'
return stcs
def induced_power(epochs, Fs, frequencies, use_fft=True, n_cycles=7,
n_jobs=1):
"""Compute time induced power and inter-trial phase-locking factor
The time frequency decomposition is done with Morlet wavelets
Parameters
----------
epochs : array
3D array of shape [n_epochs, n_channels, n_times]
Fs : float
sampling Frequency
frequencies : array
Array of frequencies of interest
use_fft : bool
Compute transform with fft based convolutions or temporal
convolutions.
n_cycles : int
The number of cycles in the wavelet
n_jobs : int
The number of CPUs used in parallel. All CPUs are used in -1.
Requires joblib package.
Returns
-------
power : 2D array
Induced power (Channels x Frequencies x Timepoints).
Squared amplitude of time-frequency coefficients.
phase_lock : 2D array
Phase locking factor in [0, 1] (Channels x Frequencies x Timepoints)
"""
n_frequencies = len(frequencies)
n_epochs, n_channels, n_times = epochs.shape
# Precompute wavelets for given frequency range to save time
Ws = morlet(Fs, frequencies, n_cycles=n_cycles)
try:
import joblib
except ImportError:
print "joblib not installed. Cannot run in parallel."
n_jobs = 1
if n_jobs == 1:
psd = np.empty((n_channels, n_frequencies, n_times))
plf = np.empty((n_channels, n_frequencies, n_times), dtype=np.complex)
for c in range(n_channels):
X = np.squeeze(epochs[:, c, :])
psd[c], plf[c] = _time_frequency(X, Ws, use_fft)
else:
from joblib import Parallel, delayed
psd_plf = Parallel(n_jobs=n_jobs)(
delayed(_time_frequency)(
np.squeeze(epochs[:, c, :]), Ws, use_fft)
for c in range(n_channels))
psd = np.zeros((n_channels, n_frequencies, n_times))
plf = np.zeros((n_channels, n_frequencies, n_times), dtype=np.complex)
for c, (psd_c, plf_c) in enumerate(psd_plf):
psd[c, :, :], plf[c, :, :] = psd_c, plf_c
psd /= n_epochs
plf = np.abs(plf) / n_epochs
return psd, plf
def single_trial_power(epochs, Fs, frequencies, use_fft=True, n_cycles=7,
baseline=None, baseline_mode='ratio', times=None,
n_jobs=1):
"""Compute time-frequency power on single epochs
Parameters
----------
epochs : instance Epochs | array of shape [n_epochs, n_channels, n_times]
The epochs
Fs : float
Sampling rate
frequencies : array-like
The frequencies
use_fft : bool
Use the FFT for convolutions or not.
n_cycles : float
The number of cycles in the Morlet wavelet
baseline: None (default) or tuple of length 2
The time interval to apply baseline correction.
If None do not apply it. If baseline is (a, b)
the interval is between "a (s)" and "b (s)".
If a is None the beginning of the data is used
and if b is None then b is set to the end of the interval.
If baseline is equal ot (None, None) all the time
interval is used.
baseline_mode : None | 'ratio' | 'zscore'
Do baseline correction with ratio (power is divided by mean
power during baseline) or zscore (power is divided by standard
deviatio of power during baseline after substracting the mean,
power = [power - mean(power_baseline)] / std(power_baseline))
times : array
Required to define baseline
n_jobs : int
The number of epochs to process at the same time
Returns
-------
power : list of 2D array
Each element of the list the the power estimate for an epoch.
"""
mode = 'same'
n_frequencies = len(frequencies)
n_epochs, n_channels, n_times = epochs.shape
# Precompute wavelets for given frequency range to save time
Ws = morlet(Fs, frequencies, n_cycles=n_cycles)
try:
from scikits.learn.externals.joblib import Parallel, delayed
parallel = Parallel(n_jobs)
my_cwt = delayed(cwt)
except ImportError:
print "joblib not installed. Cannot run in parallel."
n_jobs = 1
my_cwt = cwt
parallel = list
print "Computing time-frequency power on single epochs..."
power = np.empty((n_epochs, n_channels, n_frequencies, n_times),
dtype=np.float)
if n_jobs == 1:
for k, e in enumerate(epochs):
power[k] = np.abs(cwt(e, Ws, mode)) ** 2
else:
# Precompute tf decompositions in parallel
tfrs = parallel(my_cwt(e, Ws, use_fft, mode) for e in epochs)
for k, tfr in enumerate(tfrs):
power[k] = np.abs(tfr) ** 2
# Run baseline correction
if baseline is not None:
if times is None:
raise ValueError('times parameter is required to define baseline')
print "Applying baseline correction ..."
bmin = baseline[0]
bmax = baseline[1]
if bmin is None:
imin = 0
else:
imin = int(np.where(times >= bmin)[0][0])
if bmax is None:
imax = len(times)
else:
imax = int(np.where(times <= bmax)[0][-1]) + 1
mean_baseline_power = np.mean(power[:, :, :, imin:imax], axis=3)
if baseline_mode is 'ratio':
power /= mean_baseline_power[:, :, :, None]
elif baseline_mode is 'zscore':
power -= mean_baseline_power[:, :, :, None]
power /= np.std(power[:, :, :, imin:imax], axis=3)[:, :, :, None]
else:
print "No baseline correction applied..."
return power
