def get_csp_filters_multiclass(self, ClassData, n=5, csp_type="RG"):
C = np.zeros((len(ClassData), ClassData[0].shape[1], ClassData[1].shape[1]))
sample_weights = list()
numClasses = len(ClassData)
numChannels = ClassData[0].shape[1]
for i in range(numClasses):
class_ = np.transpose(ClassData[i], [1, 0, 2])
class_ = class_.reshape(numChannels, -1)
C[i] = np.cov(class_)
weight = ClassData[i].shape[1]
sample_weights.append(weight)
w, v = _ajd_pham(C)
mean_cov = np.average(C, axis=0, weights=sample_weights)
w_, v_ = np.linalg.eig(C[0])
w = np.asarray(w)
max_filters = []
min_filters = []
if csp_type == "STD":
w_sort = np.sort(w)
max_filters = [v[:, np.where(w == k)[0][0]] for k in w_sort[-n:]]
min_filters = [v[:, np.where(w == k)[0][0]] for k in w_sort[:n]]
elif csp_type == "RG":
max_div = []
min_div = []
for ind, k in enumerate(w):
vect = v[:, ind]
div = np.dot(np.dot(np.transpose(vect), R1), vect) / np.dot(np.dot(np.transpose(vect), R2), vect)
max_div.append(div)
min_div.append(div)
div_idx = len(max_div) - 1
while max_div[div_idx] > max_div[div_idx - 1] and div_idx > 0:
t = max_div[div_idx - 1]
max_div[div_idx - 1] = max_div[div_idx]
max_div[div_idx] = t
div_idx -= 1
max_filters.insert(div_idx, vect)
if len(max_div) > n:
max_div = max_div[:n]
max_filters = max_filters[:n]
div_idx = len(min_div) - 1
while min_div[div_idx] < min_div[div_idx - 1] and div_idx > 0:
t = min_div[div_idx - 1]
min_div[div_idx - 1] = min_div[div_idx]
min_div[div_idx] = t
div_idx -= 1
min_filters.insert(div_idx, vect)
if len(min_div) > n:
min_div = min_div[:n]
min_filters = min_filters[:n]
print('max', max_div)
print('min', min_div)
self.csp_filters = min_filters + max_filters
def test_ajd():
"""Test approximate joint diagonalization."""
# The implementation shuold obtain the same
# results as the Matlab implementation by Pham Dinh-Tuan.
# Generate a set of cavariances matrices for test purpose
n_times, n_channels = 10, 3
seed = np.random.RandomState(0)
diags = 2.0 + 0.1 * seed.randn(n_times, n_channels)
A = 2 * seed.rand(n_channels, n_channels) - 1
A /= np.atleast_2d(np.sqrt(np.sum(A**2, 1))).T
covmats = np.empty((n_times, n_channels, n_channels))
for i in range(n_times):
covmats[i] = np.dot(np.dot(A, np.diag(diags[i])), A.T)
V, D = _ajd_pham(covmats)
# Results obtained with original matlab implementation
V_matlab = [[-3.507280775058041, -5.498189967306344, 7.720624541198574],
[0.694689013234610, 0.775690358505945, -1.162043086446043],
[-0.592603135588066, -0.598996925696260, 1.009550086271192]]
assert_array_almost_equal(V, V_matlab)
def test_ajd():
"""Test if Approximate joint diagonalization implementation obtains same
results as the Matlab implementation by Pham Dinh-Tuan.
"""
# Generate a set of cavariances matrices for test purpose
n_times, n_channels = 10, 3
seed = np.random.RandomState(0)
diags = 2.0 + 0.1 * seed.randn(n_times, n_channels)
A = 2 * seed.rand(n_channels, n_channels) - 1
A /= np.atleast_2d(np.sqrt(np.sum(A ** 2, 1))).T
covmats = np.empty((n_times, n_channels, n_channels))
for i in range(n_times):
covmats[i] = np.dot(np.dot(A, np.diag(diags[i])), A.T)
V, D = _ajd_pham(covmats)
# Results obtained with original matlab implementation
V_matlab = [[-3.507280775058041, -5.498189967306344, 7.720624541198574],
[0.694689013234610, 0.775690358505945, -1.162043086446043],
[-0.592603135588066, -0.598996925696260, 1.009550086271192]]
assert_array_almost_equal(V, V_matlab)