def _compute_rank(p, subj, run_indices):
"""Compute rank of the data."""
epochs_fnames, _ = get_epochs_evokeds_fnames(p, subj, p.analyses)
_, fif_file = epochs_fnames
epochs = read_epochs(fif_file) # .crop(p.bmin, p.bmax) maybe someday...?
meg, eeg = 'meg' in epochs, 'eeg' in epochs
rank = dict()
epochs.apply_proj()
if p.cov_rank_method == 'estimate_rank':
# old way
if meg:
eps = epochs.copy().pick_types(meg=meg, eeg=False)
eps = eps.get_data().transpose([1, 0, 2])
eps = eps.reshape(len(eps), -1)
if 'grad' in epochs and 'mag' in epochs: # Neuromag
key = 'meg'
else:
key = 'grad' if 'grad' in epochs else 'mag'
rank[key] = estimate_rank(eps, tol=p.cov_rank_tol)
if eeg:
eps = epochs.copy().pick_types(meg=False, eeg=eeg)
eps = eps.get_data().transpose([1, 0, 2])
eps = eps.reshape(len(eps), -1)
rank['eeg'] = estimate_rank(eps, tol=p.cov_rank_tol)
else:
assert p.cov_rank_method == 'compute_rank'
# new way
rank = compute_rank(epochs, tol=p.cov_rank_tol, tol_kind='relative')
for k, v in rank.items():
print(' : %s rank %2d' % (k.upper(), v), end='')
return rank
def test_estimate_rank():
"""Test rank estimation."""
data = np.eye(10)
assert_array_equal(
estimate_rank(data, return_singular=True)[1], np.ones(10))
data[0, 0] = 0
assert estimate_rank(data) == 9
pytest.raises(ValueError, estimate_rank, data, 'foo')
def test_estimate_rank():
"""Test rank estimation."""
data = np.eye(10)
assert_array_equal(estimate_rank(data, return_singular=True)[1],
np.ones(10))
data[0, 0] = 0
assert estimate_rank(data) == 9
pytest.raises(ValueError, estimate_rank, data, 'foo')
def check_rank_cov_matrix(cov_mat, epochs):
"""Check the covariance matrix and report rank and channel number."""
ch_num = len(cov_mat.ch_names)
rank_cov = estimate_rank(cov_mat['data'], tol='auto')
cov_ok = ch_num - rank_cov
cond_num = cond(cov_mat['data'])
if cov_ok > 0 and cond_num > 10e15:
# cond num: amazing is < 10e4, really shitty is > 10e12, horrible 10e19
print_ok = 'WARNING! Bad cov matrix.'
else:
print_ok = 'ok.'
print('%i trials, %i channels, cov matrix has rank %i and condition '
'number %s: %s' % (len(epochs.events), ch_num, rank_cov,
'{:.2E}'.format(Decimal(cond_num)), print_ok))
def _make_diagonal_noise_matrix(csd, reg):
"""Make a diagonal matrix suitable for using as a noise CSD.
Starts with an identity matrix and scales it to the smallest singular value
of the CSD matrix that is comfortably larger than 0.
Parameters
----------
csd : ndarray, shape (n_series, n_series)
The data cross-spectral density (CSD) matrix.
reg : float
The regulariation parameter used when inverting the CSD matrix.
Returns
-------
noise : ndarray, shape (n_series, n_series)
A suitable noise cross-spectral density (CSD) matrix.
"""
rank, s = estimate_rank(csd, tol='auto', norm=False, return_singular=True)
last_singular_value = abs(s[rank - 1])
ratio = max(abs(reg), last_singular_value)
return np.eye(len(csd)) * ratio
