def _compute_node_covariance(self, small_ts, input_ts_h5): """ Compute the temporal covariance between nodes in a TimeSeries dataType. A nodes x nodes matrix is returned for each (state-variable, mode). """ data_shape = small_ts.data.shape # (nodes, nodes, state-variables, modes) result_shape = (data_shape[2], data_shape[2], data_shape[1], data_shape[3]) self.log.info("result shape will be: %s" % str(result_shape)) result = numpy.zeros(result_shape) # One inter-node temporal covariance matrix for each state-var & mode. for mode in range(data_shape[3]): for var in range(data_shape[1]): data = input_ts_h5.data[:, var, :, mode] data = data - data.mean(axis=0)[numpy.newaxis, 0] result[:, :, var, mode] = numpy.cov(data.T) self.log.debug("result") self.log.debug(narray_describe(result)) covariance = Covariance(source=small_ts, array_data=result) return covariance
def _compute_cross_correlation(self, small_ts, input_ts_h5): """ Cross-correlate two one-dimensional arrays. Return a CrossCorrelation datatype with result. """ # (tpts, nodes, nodes, state-variables, modes) result_shape = self._result_shape(small_ts.data.shape) self.log.info("result shape will be: %s" % str(result_shape)) result = numpy.zeros(result_shape) # TODO: For region level, 4s, 2000Hz, this takes ~3hours...(which makes node_coherence seem positively speedy # Probably best to add a keyword for offsets, so we just compute +- some "small" range... # One inter-node correlation, across offsets, for each state-var & mode. for mode in range(result_shape[4]): for var in range(result_shape[3]): data = input_ts_h5.data[:, var, :, mode] data = data - data.mean(axis=0)[numpy.newaxis, :] # TODO: Work out a way around the 4 level loop: for n1 in range(result_shape[1]): for n2 in range(result_shape[2]): result[:, n1, n2, var, mode] = correlate(data[:, n1], data[:, n2], mode="same") self.log.debug("result") self.log.debug(narray_describe(result)) offset = (small_ts.sample_period * numpy.arange(-numpy.floor(result_shape[0] / 2.0), numpy.ceil(result_shape[0] / 2.0))) cross_corr = CrossCorrelation(source=small_ts, array_data=result, time=offset) return cross_corr
def _compute_correlation_coefficients(self, ts_h5, t_start, t_end): """ Compute the correlation coefficients of a 2D array (tpts x nodes). Yields an array of size nodes x nodes x state-variables x modes. The time interval over which the correlation coefficients are computed is defined by t_start, t_end See: http://docs.scipy.org/doc/numpy/reference/generated/numpy.corrcoef.html """ # (nodes, nodes, state-variables, modes) input_shape = ts_h5.data.shape result_shape = self._result_shape(input_shape) self.log.info("result shape will be: %s" % str(result_shape)) result = numpy.zeros(result_shape) t_lo = int((1. / self.input_time_series_index.sample_period) * (t_start - self.input_time_series_index.sample_period)) t_hi = int((1. / self.input_time_series_index.sample_period) * (t_end - self.input_time_series_index.sample_period)) t_lo = max(t_lo, 0) t_hi = max(t_hi, input_shape[0]) # One correlation coeff matrix, for each state-var & mode. for mode in range(result_shape[3]): for var in range(result_shape[2]): current_slice = tuple([ slice(t_lo, t_hi + 1), slice(var, var + 1), slice(input_shape[2]), slice(mode, mode + 1) ]) data = ts_h5.data[current_slice].squeeze() result[:, :, var, mode] = numpy.corrcoef(data.T) self.log.debug("result") self.log.debug(narray_describe(result)) return result
def calculate_complex_cross_coherence(time_series, epoch_length, segment_length, segment_shift, window_function, average_segments, subtract_epoch_average, zeropad, detrend_ts, max_freq, npat): """ # type: (TimeSeries, float, float, float, str, bool, bool, int, bool, float, float) -> ComplexCoherenceSpectrum Calculate the FFT, Cross Coherence and Complex Coherence of time_series broken into (possibly) epochs and segments of length `epoch_length` and `segment_length` respectively, filtered by `window_function`. Parameters __________ time_series : TimeSeries The timeseries for which the CrossCoherence and ComplexCoherence is to be computed. epoch_length : float In general for lengthy EEG recordings (~30 min), the timeseries are divided into equally sized segments (~ 20-40s). These contain the event that is to be characterized by means of the cross coherence. Additionally each epoch block will be further divided into segments to which the FFT will be applied. segment_length : float The segment length determines the frequency resolution of the resulting power spectra -- longer windows produce finer frequency resolution. segment_shift : float Time length by which neighboring segments are shifted. e.g. `segment shift` = `segment_length` / 2 means 50% overlapping segments. window_function : str Windowing functions can be applied before the FFT is performed. average_segments : bool Flag. If `True`, compute the mean Cross Spectrum across segments. subtract_epoch_average: bool Flag. If `True` and if the number of epochs is > 1, you can optionally subtract the mean across epochs before computing the complex coherence. zeropad : int Adds `n` zeros at the end of each segment and at the end of window_function. It is not yet functional. detrend_ts : bool Flag. If `True` removes linear trend along the time dimension before applying FFT. max_freq : float Maximum frequency points (e.g. 32., 64., 128.) represented in the output. Default is segment_length / 2 + 1. npat : float This attribute appears to be related to an input projection matrix... Which is not yet implemented. """ # self.time_series.trait["data"].log_debug(owner=cls_attr_name) tpts = time_series.data.shape[0] time_series_length = tpts * time_series.sample_period if len(time_series.data.shape) > 2: time_series_data = numpy.squeeze((time_series.data.mean(axis=-1)).mean(axis=1)) # Divide time-series into epochs, no overlapping if epoch_length > 0.0: nepochs = int(numpy.floor(time_series_length / epoch_length)) epoch_tpts = int(epoch_length / time_series.sample_period) time_series_length = epoch_length tpts = epoch_tpts else: epoch_length = time_series_length nepochs = int(numpy.ceil(time_series_length / epoch_length)) # Segment time-series, overlapping if necessary nseg = int(numpy.floor(time_series_length / segment_length)) if nseg > 1: seg_tpts = int(segment_length / time_series.sample_period) seg_shift_tpts = int(segment_shift / time_series.sample_period) nseg = int(numpy.floor((tpts - seg_tpts) / seg_shift_tpts) + 1) else: segment_length = time_series_length seg_tpts = time_series_data.shape[0] # Frequency nfreq = int(numpy.min([max_freq, numpy.floor((seg_tpts + zeropad) / 2.0) + 1])) resulted_shape, av_result_shape = complex_coherence_result_shape(time_series.data.shape, max_freq, epoch_length, segment_length, segment_shift, time_series.sample_period, zeropad, average_segments) cs = numpy.zeros(resulted_shape, dtype=numpy.complex128) av = numpy.matrix(numpy.zeros(av_result_shape, dtype=numpy.complex128)) coh = numpy.zeros(resulted_shape, dtype=numpy.complex128) # Apply windowing function if window_function is not None: if window_function not in SUPPORTED_WINDOWING_FUNCTIONS: log.error("Windowing function is: %s" % window_function) log.error("Must be in: %s" % str(SUPPORTED_WINDOWING_FUNCTIONS)) window_func = eval("".join(("numpy.", window_function))) win = window_func(seg_tpts) window_mask = (numpy.kron(numpy.ones((time_series_data.shape[1], 1)), win)).T nave = 0 for j in numpy.arange(nepochs): data = time_series_data[j * epoch_tpts:(j + 1) * epoch_tpts, :] for i in numpy.arange(nseg): # average over all segments; ts = data[i * seg_shift_tpts: i * seg_shift_tpts + seg_tpts, :] if detrend_ts: ts = sp_signal.detrend(ts, axis=0) datalocfft = numpy.fft.fft(ts * window_mask, axis=0) datalocfft = numpy.matrix(datalocfft) for f in numpy.arange(nfreq): # for all frequencies if npat == 1: if not average_segments: cs[:, :, f, i] += numpy.conjugate(datalocfft[f, :].conj().T * datalocfft[f, :]) av[:, f, i] += numpy.conjugate(datalocfft[f, :].conj().T) else: cs[:, :, f] += numpy.conjugate(datalocfft[f, :].conj().T * datalocfft[f, :]) av[:, f] += numpy.conjugate(datalocfft[f, :].conj().T) else: if not average_segments: cs[:, :, f, j, i] = numpy.conjugate(datalocfft[f, :].conj().T * datalocfft[f, :]) av[:, f, j, i] = numpy.conjugate(datalocfft[f, :].conj().T) else: cs[:, :, f, j] += numpy.conjugate(datalocfft[f, :].conj().T * datalocfft[f, :]) av[:, f, j] += numpy.conjugate(datalocfft[f, :].conj().T) del datalocfft nave += 1.0 # End of FORs if not average_segments: cs = cs / nave av = av / nave else: nave = nave * nseg cs = cs / nave av = av / nave # Subtract average for f in numpy.arange(nfreq): if subtract_epoch_average: if npat == 1: if not average_segments: for i in numpy.arange(nseg): cs[:, :, f, i] = cs[:, :, f, i] - av[:, f, i] * av[:, f, i].conj().T else: cs[:, :, f] = cs[:, :, f] - av[:, f] * av[:, f].conj().T else: if not average_segments: for i in numpy.arange(nseg): for j in numpy.arange(nepochs): cs[:, :, f, j, i] = cs[:, :, f, j, i] - av[:, f, j, i] * av[:, f, j, i].conj().T else: for j in numpy.arange(nepochs): cs[:, :, f, j] = cs[:, :, f, j] - av[:, f, j] * av[:, f, j].conj().T # Compute Complex Coherence ndim = len(cs.shape) if ndim == 3: for i in numpy.arange(cs.shape[2]): temp = numpy.matrix(cs[:, :, i]) coh[:, :, i] = cs[:, :, i] / numpy.sqrt(temp.diagonal().conj().T * temp.diagonal()) elif ndim == 4: for i in numpy.arange(cs.shape[2]): for j in numpy.arange(cs.shape[3]): temp = numpy.matrix(numpy.squeeze(cs[:, :, i, j])) coh[:, :, i, j] = temp / numpy.sqrt(temp.diagonal().conj().T * temp.diagonal().T) log.debug("result") log.debug(narray_describe(cs)) spectra = spectral.ComplexCoherenceSpectrum(source=time_series, array_data=coh, cross_spectrum=cs, epoch_length=epoch_length, segment_length=segment_length, windowing_function=window_function) return spectra
def compute_fast_fourier_transform(time_series, segment_length, window_function, detrend): """ # type: (TimeSeries, float, function, bool) -> FourierSpectrum Calculate the FFT of time_series broken into segments of length segment_length and filtered by window_function. Parameters __________ time_series : TimeSeries The TimeSeries to which the FFT is to be applied. segment_length : float The segment length determines the frequency resolution of the resulting power spectra -- longer windows produce finer frequency resolution window_function : str Windowing functions can be applied before the FFT is performed. Default is None, possibilities are: 'hamming'; 'bartlett';'blackman'; and 'hanning'. See, numpy.<function_name>. detrend : bool Default is True, False means no detrending is performed on the time series. """ tpts = time_series.data.shape[0] time_series_length = tpts * time_series.sample_period # Segment time-series, overlapping if necessary nseg = int(numpy.ceil(time_series_length / segment_length)) if nseg > 1: seg_tpts = numpy.ceil(segment_length / time_series.sample_period) overlap = (seg_tpts * nseg - tpts) / (nseg - 1.0) starts = [max(seg * (seg_tpts - overlap), 0) for seg in range(nseg)] segments = [ time_series.data[int(start):int(start) + int(seg_tpts)] for start in starts ] segments = [segment[:, :, :, :, numpy.newaxis] for segment in segments] ts = numpy.concatenate(segments, axis=4) else: segment_length = time_series_length ts = time_series.data[:, :, :, :, numpy.newaxis] seg_tpts = ts.shape[0] log.debug("Segment length being used is: %s" % segment_length) # Base-line correct the segmented time-series if detrend: ts = scipy.signal.detrend(ts, axis=0) log.debug("time_series " + narray_describe(ts)) # Apply windowing function if window_function is not None: wf = SUPPORTED_WINDOWING_FUNCTIONS[window_function] window_mask = numpy.reshape(wf(int(seg_tpts)), (int(seg_tpts), 1, 1, 1, 1)) ts = ts * window_mask # Calculate the FFT result = numpy.fft.fft(ts, axis=0) nfreq = result.shape[0] // 2 result = result[1:nfreq + 1, :] log.debug("result " + narray_describe(result)) spectra = FourierSpectrum(source=time_series, segment_length=segment_length, array_data=result, windowing_function=window_function) spectra.configure() return spectra
def _compute_fcd_matrix(self, ts_h5): self.log.debug("timeseries_h5.data") self.log.debug(narray_describe(ts_h5.data[:])) input_shape = ts_h5.data.shape result_shape = self._result_shape(input_shape) fcd = np.zeros(result_shape) fc_stream = { } # dict where the fc calculated over the sliding window will be stored for mode in range(result_shape[3]): for var in range(result_shape[2]): start = -self.actual_sp # in order to well initialize the first starting point of the FC stream for nfcd in range(result_shape[0]): start += self.actual_sp current_slice = tuple([ slice(int(start), int(start + self.actual_sw) + 1), slice(var, var + 1), slice(input_shape[2]), slice(mode, mode + 1) ]) data = ts_h5.read_data_slice(current_slice).squeeze() fc = np.corrcoef(data.T) # the triangular part of the fc is organized as a vector, excluding the diagonal (always ones) triangular = np.triu_indices(len(fc), 1) fc_stream[nfcd] = fc[triangular] for i in range(result_shape[0]): j = i while j < result_shape[0]: fci = fc_stream[i] fcj = fc_stream[j] fcd[i, j, var, mode] = np.corrcoef(fci, fcj)[0, 1] fcd[j, i, var, mode] = fcd[i, j, var, mode] j += 1 self.log.debug("FCD") self.log.debug(narray_describe(fcd)) num_eig = 3 # number of the eigenvector that will be extracted eigvect_dict = { } # holds eigenvectors of the fcs calculated over the epochs, key1=mode, key2=var, key3=numb ep eigval_dict = { } # holds eigenvalues of the fcs calculated over the epochs, key1=mode, key2=var, key3=numb ep fcd_segmented = None for mode in range(result_shape[3]): eigvect_dict[mode] = {} eigval_dict[mode] = {} for var in range(result_shape[2]): eigvect_dict[mode][var] = {} eigval_dict[mode][var] = {} fcd_matrix = fcd[:, :, var, mode] [xir, xir_cutoff] = self._spectral_embedding(fcd_matrix) epochs_extremes = self._epochs_interval( xir, xir_cutoff, self.actual_sp, self.actual_sw) fcd_segmented = fcd.copy() if epochs_extremes.shape[0] <= 1: # means that there are no more than 1 epochs of stability, thus the eigenvectors of # the FC calculated over the entire TimeSeries will be calculated epochs_extremes = np.zeros((2, 2), dtype=float) epochs_extremes[1, 1] = input_shape[ 0] # [0,0] set in order to skip the first epoch else: # means that more than 1 epochs of stability is identified thus fcd_segmented is calculated fcd_segmented[xir > xir_cutoff, :, var, mode] = 1.1 fcd_segmented[:, xir > xir_cutoff, var, mode] = 1.1 for ep in range(1, epochs_extremes.shape[0]): eigvect_dict[mode][var][ep] = [] eigval_dict[mode][var][ep] = [] current_slice = tuple([ slice(int(epochs_extremes[ep][0]), int(epochs_extremes[ep][1]) + 1), slice(var, var + 1), slice(input_shape[2]), slice(mode, mode + 1) ]) data = ts_h5.read_data_slice(current_slice).squeeze() fc = np.corrcoef( data.T) # calculate fc over the epoch of stability eigval_matrix, eigvect_matrix = linalg.eig(fc) eigval_matrix = np.real(eigval_matrix) eigvect_matrix = np.real(eigvect_matrix) eigval_matrix = eigval_matrix / np.sum( np.abs(eigval_matrix) ) # normalize eigenvalues to [0 and 1) for en in range(num_eig): index = np.argmax(eigval_matrix) eigvect_dict[mode][var][ep].append( abs(eigvect_matrix[:, index])) eigval_dict[mode][var][ep].append(eigval_matrix[index]) eigval_matrix[index] = 0 return [fcd, fcd_segmented, eigvect_dict, eigval_dict]