def _compute_xDAWN_filters(self, X, D):
# Compute xDAWN spatial filters
# QR decompositions of X and D
if map(int, __import__("scipy").__version__.split('.')) >= [0, 9, 0]:
# NOTE: mode='economy'required since otherwise the memory
# consumption is excessive
Qx, Rx = qr(X, overwrite_a=True, mode='economic')
Qd, Rd = qr(D, overwrite_a=True, mode='economic')
else:
# NOTE: econ=True required since otherwise the memory consumption
# is excessive
Qx, Rx = qr(X, overwrite_a=True, econ=True)
Qd, Rd = qr(D, overwrite_a=True, econ=True)
# Singular value decomposition of Qd.T Qx
# NOTE: full_matrices=True required since otherwise we do not get
# num_channels filters.
Phi, Lambda, Psi = numpy.linalg.svd(numpy.dot(Qd.T, Qx),
full_matrices=True)
Psi = Psi.T
# Construct the spatial filters
for i in range(Psi.shape[1]):
# Construct spatial filter with index i as Rx^-1*Psi_i
ui = numpy.dot(numpy.linalg.inv(Rx), Psi[:, i])
if i == 0:
filters = numpy.atleast_2d(ui).T
else:
filters = numpy.hstack((filters, numpy.atleast_2d(ui).T))
return filters
def _compute_xDAWN_filters(self, X, D):
# Compute xDAWN spatial filters
# QR decompositions of X and D
if map(int, __import__("scipy").__version__.split('.')) >= [0,9,0]:
# NOTE: mode='economy'required since otherwise the memory
# consumption is excessive
Qx, Rx = qr(X, overwrite_a=True, mode='economic')
Qd, Rd = qr(D, overwrite_a=True, mode='economic')
else:
# NOTE: econ=True required since otherwise the memory consumption
# is excessive
Qx, Rx = qr(X, overwrite_a=True, econ=True)
Qd, Rd = qr(D, overwrite_a=True, econ=True)
# Singular value decomposition of Qd.T Qx
# NOTE: full_matrices=True required since otherwise we do not get
# num_channels filters.
Phi, Lambda, Psi = numpy.linalg.svd(numpy.dot(Qd.T, Qx),
full_matrices=True)
Psi = Psi.T
# Construct the spatial filters
for i in range(Psi.shape[1]):
# Construct spatial filter with index i as Rx^-1*Psi_i
ui = numpy.dot(numpy.linalg.inv(Rx), Psi[:,i])
if i == 0:
filters = numpy.atleast_2d(ui).T
else:
filters = numpy.hstack((filters, numpy.atleast_2d(ui).T))
return filters
def _stop_training(self, debug=False):
# The following if statement is needed only to account for
# different versions of scipy
if map(int, __import__("scipy").__version__.split('.')) >= [0, 9, 0]:
# NOTE: mode='economy'required since otherwise
# the memory consumption is excessive;
# QR decompositions of X
Qx, Rx = qr(self.X, overwrite_a=True, mode='economic')
# QR decompositions of D
Qd, Rd = qr(self.D, overwrite_a=True, mode='economic')
else:
# NOTE: econ=True required since otherwise
# the memory consumption is excessive
# QR decompositions of X
Qx, Rx = qr(self.X, overwrite_a=True, econ=True)
# QR decompositions of D
Qd, Rd = qr(self.D, overwrite_a=True, econ=True)
# self.X = None # Free memory
# self.D = None # Free memory
# Singular value decomposition of Qd.T Qx
# NOTE: full_matrices=True required since otherwise we do not get
# num_channels filters.
self.Phi, self.Lambda, self.Psi = \
numpy.linalg.svd(numpy.dot(Qd.T, Qx), full_matrices=True)
self.Psi = self.Psi.T
SNR = numpy.zeros(self.X.shape[1])
# Construct the spatial filters
for i in range(self.Psi.shape[1]):
# Construct spatial filter with index i as Rx^-1*Psi_i
ui = numpy.dot(numpy.linalg.inv(Rx), self.Psi[:, i])
wi = numpy.dot(Rx.T, self.Psi[:, i])
if i < self.Phi.shape[1]:
ai = numpy.dot(numpy.dot(numpy.linalg.inv(Rd), self.Phi[:, i]),
self.Lambda[i])
if i == 0:
self.filters = numpy.atleast_2d(ui).T
self.wi = numpy.atleast_2d(wi)
self.ai = numpy.atleast_2d(ai)
else:
self.filters = numpy.hstack(
(self.filters, numpy.atleast_2d(ui).T))
self.wi = numpy.vstack((self.wi, numpy.atleast_2d(wi)))
if i < self.Phi.shape[1]:
self.ai = numpy.vstack((self.ai, numpy.atleast_2d(ai)))
a = numpy.dot(self.D, ai.T)
b = numpy.dot(self.X, ui)
# b.view(numpy.ndarray)
# bb = numpy.dot(b.T, b)
# aa = numpy.dot(a.T, a)
SNR[i] = numpy.dot(a.T, a) / numpy.dot(b.T, b)
self.SNR = SNR
self.D = None
self.X = None
def _stop_training(self, debug=False):
# The following if statement is needed only to account for
# different versions of scipy
if map(int, __import__("scipy").__version__.split('.')) >= [0, 9, 0]:
# NOTE: mode='economy'required since otherwise
# the memory consumption is excessive;
# QR decompositions of X
Qx, Rx = qr(self.X, overwrite_a=True, mode='economic')
# QR decompositions of D
Qd, Rd = qr(self.D, overwrite_a=True, mode='economic')
else:
# NOTE: econ=True required since otherwise
# the memory consumption is excessive
# QR decompositions of X
Qx, Rx = qr(self.X, overwrite_a=True, econ=True)
# QR decompositions of D
Qd, Rd = qr(self.D, overwrite_a=True, econ=True)
# self.X = None # Free memory
# self.D = None # Free memory
# Singular value decomposition of Qd.T Qx
# NOTE: full_matrices=True required since otherwise we do not get
# num_channels filters.
self.Phi, self.Lambda, self.Psi = \
numpy.linalg.svd(numpy.dot(Qd.T, Qx), full_matrices=True)
self.Psi = self.Psi.T
SNR = numpy.zeros(self.X.shape[1])
# Construct the spatial filters
for i in range(self.Psi.shape[1]):
# Construct spatial filter with index i as Rx^-1*Psi_i
ui = numpy.dot(numpy.linalg.inv(Rx), self.Psi[:,i])
wi = numpy.dot(Rx.T, self.Psi[:,i])
if i < self.Phi.shape[1]:
ai = numpy.dot(numpy.dot(numpy.linalg.inv(Rd), self.Phi[:,i]),
self.Lambda[i])
if i == 0:
self.filters = numpy.atleast_2d(ui).T
self.wi = numpy.atleast_2d(wi)
self.ai = numpy.atleast_2d(ai)
else:
self.filters = numpy.hstack((self.filters,
numpy.atleast_2d(ui).T))
self.wi = numpy.vstack((self.wi, numpy.atleast_2d(wi)))
if i < self.Phi.shape[1]:
self.ai = numpy.vstack((self.ai, numpy.atleast_2d(ai)))
a = numpy.dot(self.D, ai.T)
b = numpy.dot(self.X, ui)
# b.view(numpy.ndarray)
# bb = numpy.dot(b.T, b)
# aa = numpy.dot(a.T, a)
SNR[i] = numpy.dot(a.T, a)/numpy.dot(b.T, b)
self.SNR = SNR
self.D = None
self.X = None