def SPD_Project(mat):
# force symmetric
mat = (mat + mat.T) / 2.0
eig, eigv = linalg.eig(mat)
eig = T.maximum(eig, 0)
eig = T.diag(eig)
return eigv.dot(eig).dot(eigv.T)
def th_MvLDAN(data_inputs, labels):
n_view = len(data_inputs)
dtype = 'float32'
mean = []
std = []
data = []
for v in range(n_view):
_data = theano.shared(data_inputs[v])
_mean = T.mean(_data, axis=0).reshape([1, -1])
_std = T.std(_data, axis=0).reshape([1, -1])
_std += T.eq(_std, 0).astype(dtype)
data.append((_data - _mean) / _std)
mean.append(_mean)
std.append(_std)
Sw, Sb, _ = th_MvLDAN_Sw_Sb(data, labels)
from theano.tensor import nlinalg
eigvals, eigvecs = nlinalg.eig(T.dot(nlinalg.matrix_inverse(Sw), Sb))
# evals = slinalg.eigvalsh(Sb, Sw)
mean = list(theano.function([], mean)())
std = list(theano.function([], std)())
eigvals, eigvecs = theano.function([], [eigvals, eigvecs])()
inx = np.argsort(eigvals)[::-1]
eigvals = eigvals[inx]
eigvecs = eigvecs[:, inx]
W = []
pre = 0
for v in range(n_view):
W.append(eigvecs[pre:pre + mean[v].shape[1], :])
pre += mean[v].shape[1]
return [mean, std], W, eigvals
def SPD_Project(mat):
# force symmetric
mat = (mat+mat.T)/2.0
eig, eigv = linalg.eig(mat)
eig = T.maximum(eig, 0)
eig = T.diag(eig)
return eigv.dot(eig).dot(eigv.T)
def __theano_SPDproject(self, mat):
# force symmetric
mat = (mat + mat.T) / 2.0
eig, eigv = linalg.eig(mat)
eig = T.maximum(eig, 0)
eig = T.diag(eig)
return eigv.dot(eig).dot(eigv.T)
def __theano_SPDproject(self, mat):
# force symmetric
mat = (mat+mat.T)/2.0
eig, eigv = linalg.eig(mat)
eig = T.maximum(eig, 0)
eig = T.diag(eig)
return eigv.dot(eig).dot(eigv.T)
def PSD_Project(mat):
mat = (mat+mat.T)/2.0
eig, eigv = linalg.eig(mat)
eig = T.maximum(eig, 0)
eig = T.diag(eig)
return eigv.dot(eig).dot(eigv.T)
def _project_sd(self, mat):
eig, eigv = linalg.eig(mat)
eig = T.diag(T.maximum(eig, 0))
return eigv.dot(eig).dot(linalg.matrix_inverse(eigv))
