def do_mvar_evaluation(X, morder, whit_max=3., whit_min=1., thr_cons=0.8): ''' Fit MVAR model to data using scot and do some basic checks. X: array (trials, channels, samples) morder: the model order Returns: (is_white, consistency, is_stable) ''' print('starting checks and MVAR fitting...') # tsdata_to_var from MVGC requires sources x samples x trials # X is of shape trials x sources x samples (which is what ScoT uses) X_trans = X.transpose(1, 2, 0) A, SIG, E = _tsdata_to_var(X_trans, morder) del A, SIG whi = False dw, pval = dw_whiteness(X_trans, E) if np.all(dw < whit_max) and np.all(dw > whit_min): whi = True cons = consistency(X_trans, E) del dw, pval, E from scot.var import VAR mvar = VAR(morder) mvar.fit(X) # scot func which requires shape trials x sources x samples is_st = mvar.is_stable() if cons < thr_cons or is_st is False or whi is False: print('ERROR: Model order not ideal - check parameters !!') return str(whi), cons, str(is_st)
def compute_order_extended(X, m_max, m_min=1, m_step=1, n_jobs=None, verbose=True): """ Estimate VAR order with the Bayesian Information Criterion (BIC). Parameters: ----------- X : ndarray, shape (trials, n_channels, n_samples) m_max : int The maximum model order to test, m_min : int The minimum model order to test. m_step : int The step size for checking the model order interval given by m_min and m_max. n_jobs : None | int, optional Number of jobs to run in parallel for various tasks (e.g. whiteness testing). If set to None, joblib is not used at all. Note that the main script must be guarded with `if __name__ == '__main__':` when using parallelization. verbose : bool Plot results for other information criteria as well. Returns: -------- o_m : int Estimated order using BIC2. morders : np.array of shape ((m_max - m_min) / m_step, ) The model orders corresponding to the entries in the following results arrays. ics : np.array of shape (n_ics, (m_max - m_min) / m_step) The information criteria for the different model orders. [AIC1, BIC1, AIC2, BIC2, lnFPE, HQIC]] stability : np.array of shape ((m_max - m_min) / m_step), ) Indicates if MVAR model describes stable process (covariance stationary). p_white_scot : np.array of shape ((m_max - m_min) / m_step), ) p-value that the residuals are white based on the Li-McLeod Portmanteau test implemented in SCoT. Reject hypothesis of white residuals if p is smaller than the critical p-value. p_white_dw : np.array of shape ((m_max - m_min) / m_step), n_rois) Uncorrected p-values that the residuals are white based on the Durbin-Watson test as implemented by Barnett and Seth (2012). Reject hypothesis of white residuals if all p's are smaller than the critical p-value. dw : np.array of shape ((m_max - m_min) / m_step), n_rois) The Durbin-Watson statistics. consistency : np.array of shape ((m_max - m_min) / m_step), ) Results of the MVAR consistency estimation. References: ----------- [1] provides the equation:BIC(m) = 2*log[det(Σ)]+ 2*(p**2)*m*log(N*n*m)/(N*n*m), Σ is the noise covariance matrix, p is the channels, N is the trials, n is the n_samples, m is model order. [1] Mingzhou Ding, Yonghong Chen (2008). "Granger Causality: Basic Theory and Application to Neuroscience." Elsevier Science [2] Nicoletta Nicolaou and Julius Georgiou (2013). “Autoregressive Model Order Estimation Criteria for Monitoring Awareness during Anaesthesia.” IFIP Advances in Information and Communication Technology 412 [3] Helmut Lütkepohl (2005). "New Introduction to Multiple Time Series Analysis." 1st ed. Berlin: Springer-Verlag Berlin Heidelberg. URL: https://gist.github.com/dongqunxi/b23d1679b9bffa8e458c11f93bd8d6ff """ from scot.var import VAR from scipy import linalg N, p, n = X.shape aic1 = [] bic1 = [] aic2 = [] bic2 = [] lnfpe = [] hqic = [] morders = [] stability = [] p_white_scot = [] p_white_dw = [] dw = [] consistency = [] # TODO: should this be n_total = N * n * p ??? # total number of data points: n_trials * n_samples # Esther Florin (2010): N_total is number of time points contained in each time series n_total = N * n # check model order min/max/step input if m_min >= m_max: m_min = m_max - 1 if m_min < 1: m_min = 1 if m_step < 1: m_step = 1 if m_step >= m_max: m_step = m_max for m in range(m_min, m_max + 1, m_step): morders.append(m) mvar = VAR(m, n_jobs=n_jobs) mvar.fit(X) stable = mvar.is_stable() stability.append(stable) p_white_scot_ = mvar.test_whiteness(h=m, repeats=100, get_q=False, random_state=None) white_scot_ = p_white_scot_ >= 0.05 p_white_scot.append(p_white_scot_) white_dw_, cons, dw_, pval = check_whiteness_and_consistency( X.transpose(1, 2, 0), mvar.residuals.transpose(1, 2, 0), alpha=0.05) dw.append(dw_) p_white_dw.append(pval) consistency.append(cons) sigma = mvar.rescov ######################################################################## # from [1] ######################################################################## m_aic = 2 * np.log(linalg.det(sigma)) + 2 * (p**2) * m / n_total m_bic = 2 * np.log( linalg.det(sigma)) + 2 * (p**2) * m / n_total * np.log(n_total) aic1.append(m_aic) bic1.append(m_bic) ######################################################################## # from [2] ######################################################################## m_aic2 = np.log(linalg.det(sigma)) + 2 * (p**2) * m / n_total m_bic2 = np.log( linalg.det(sigma)) + (p**2) * m / n_total * np.log(n_total) aic2.append(m_aic2) bic2.append(m_bic2) ######################################################################## # from [3] ######################################################################## # Akaike's final prediction error m_ln_fpe3 = np.log(linalg.det(sigma)) + p * np.log( (n_total + m * p + 1) / (n_total - m * p - 1)) # Hannan-Quinn criterion m_hqc3 = np.log(linalg.det(sigma)) + 2 * (p**2) * m / n_total * np.log( np.log(n_total)) lnfpe.append(m_ln_fpe3) hqic.append(m_hqc3) if verbose: results = 'Model order: ' + str(m).zfill(2) results += ' AIC1: %.2f' % m_aic results += ' BIC1: %.2f' % m_bic results += ' AIC2: %.2f' % m_aic2 results += ' BIC2: %.2f' % m_bic2 results += ' lnFPE3: %.2f' % m_ln_fpe3 results += ' HQC3: %.2f' % m_hqc3 results += ' stable: %s' % str(stable) results += ' white1: %s' % str(white_scot_) results += ' white2: %s' % str(white_dw_) results += ' DWmin: %.2f' % dw_.min() results += ' DWmax: %.2f' % dw_.max() results += ' consistency: %.4f' % cons print(results) morders = np.array(morders) o_m = morders[np.argmin(bic2)] if verbose: print('>>> Optimal model order according to BIC2 = %d' % o_m) ics = [aic1, bic1, aic2, bic2, lnfpe, hqic] ics = np.asarray(ics) stability = np.array(stability) p_white_scot = np.array(p_white_scot) p_white_dw = np.array(p_white_dw) dw = np.array(dw) consistency = np.array(consistency) return o_m, morders, ics, stability, p_white_scot, p_white_dw, dw, consistency
def check_model_order(X, p, whit_min=1.5, whit_max=2.5, check_stability=True): """ Check whiteness, consistency, and stability for all model orders k <= p. Computationally intensive but for high model orders probably faster than do_mvar_evaluation(). Parameters: ----------- X : narray, shape (n_epochs, n_sources, n_times) The data to estimate the model order for. p : int The maximum model order. whit_min : float Lower boundary for the Durbin-Watson test. whit_max : float Upper boundary for the Durbin-Watson test. check_stability : bool Check the stability condition. Time intensive since it fits a second MVAR model from scot.var.VAR Returns: -------- A: array, coefficients of the specified model SIG:array, recovariance of this model E: array, noise covariance of this model """ assert p >= 1, "The model order must be greater or equal to 1." from scot.var import VAR X_orig = X.copy() X = X.transpose(1, 2, 0) n, m, N = X.shape p1 = p + 1 q1n = p1 * n I = np.eye(n) XX = np.zeros((n, p1, m + p, N)) for k in range(p1): XX[:, k, k:k + m, :] = X AF = np.zeros((n, q1n)) AB = np.zeros((n, q1n)) k = 1 kn = k * n M = N * (m - k) kf = list(range(0, kn)) kb = list(range(q1n - kn, q1n)) XF = np.reshape(XX[:, 0:k, k:m, :], (kn, M), order='F') XB = np.reshape(XX[:, 0:k, k - 1:m - 1, :], (kn, M), order='F') CXF = np.linalg.cholesky(XF.dot(XF.T)).T CXB = np.linalg.cholesky(XB.dot(XB.T)).T AF[:, kf] = np.linalg.solve(CXF.T, I) AB[:, kb] = np.linalg.solve(CXB.T, I) del p1, XF, XB, CXF, CXB while k <= p: tempF = np.reshape(XX[:, 0:k, k:m, :], (kn, M), order='F') af = AF[:, kf] EF = af.dot(tempF) del af, tempF tempB = np.reshape(XX[:, 0:k, k - 1:m - 1, :], (kn, M), order='F') ab = AB[:, kb] EB = ab.dot(tempB) del ab, tempB CEF = np.linalg.cholesky(EF.dot(EF.T)).T CEB = np.linalg.cholesky(EB.dot(EB.T)).T R = np.dot(np.linalg.solve(CEF.T, EF.dot(EB.T)), np.linalg.inv(CEB)) del EB, CEF, CEB RF = np.linalg.cholesky(I - R.dot(R.T)).T RB = np.linalg.cholesky(I - (R.T).dot(R)).T k = k + 1 kn = k * n M = N * (m - k) kf = np.arange(kn) kb = list(range(q1n - kn, q1n)) AFPREV = AF[:, kf] ABPREV = AB[:, kb] AF[:, kf] = np.linalg.solve(RF.T, AFPREV - R.dot(ABPREV)) AB[:, kb] = np.linalg.solve(RB.T, ABPREV - R.T.dot(AFPREV)) del RF, RB, ABPREV # check MVAR model properties E = np.linalg.solve(AFPREV[:, :n], EF) E = np.reshape(E, (n, m - k + 1, N), order='F') if k > 1: whi, cons, _, _ = check_whiteness_and_consistency( X, E, whit_min, whit_max) if check_stability: mvar = VAR((k - 1)) mvar.fit( X_orig ) # scot func which requires shape trials x sources x samples is_st = mvar.is_stable() output = 'morder %d:' % (k - 1) output += ' white: %s' % str(whi) output += '; consistency: %.4f' % cons if check_stability: output += '; stable: %s' % str(is_st) print(output)