def hessian(v, p, X1, X2, Y, rowind, colind, lamb):
#P = np.dot(X,v)
#P = vecProd(X1, X2, v)
P = sparse_kronecker_multiplication_tools_python.x_gets_subset_of_A_kron_B_times_v(v, X2, X1.T, colind, rowind)
z = (1. - Y*P)
z = np.where(z>0, z, 0)
sv = np.nonzero(z)[0]
#map to rows and cols
rows = rowind[sv]
cols = colind[sv]
p_after = sparse_kronecker_multiplication_tools_python.x_gets_subset_of_A_kron_B_times_v(p, X2, X1.T, cols, rows)
p_after = sparse_kronecker_multiplication_tools_python.x_gets_A_kron_B_times_sparse_v(p_after, X1.T, X2, rows, cols)
p_after = p_after.reshape(X2.shape[1], X1.shape[1]).T.ravel()
return p_after + lamb*p
def func(v):
#REPLACE
#P = np.dot(X,v)
P = sparse_kronecker_multiplication_tools_python.x_gets_subset_of_A_kron_B_times_v(v, X2, X1.T, colind, rowind)
z = (1. - Y*P)
z = np.where(z>0, z, 0)
return np.dot(z,z)
def mv(v):
v_after = sparse_kronecker_multiplication_tools_python.x_gets_subset_of_A_kron_B_times_v(
v, X1, X2.T, label_row_inds, label_col_inds)
v_after = sparse_kronecker_multiplication_tools_python.x_gets_A_kron_B_times_sparse_v(
v_after, X1.T, X2, label_row_inds,
label_col_inds) + regparam * v
return v_after
def func(v):
#REPLACE
#P = np.dot(X,v)
P = sparse_kronecker_multiplication_tools_python.x_gets_subset_of_A_kron_B_times_v(
v, X2, X1.T, colind, rowind)
z = (1. - Y * P)
z = np.where(z > 0, z, 0)
return np.dot(z, z) + lamb * np.dot(v, v)
def hessian(v, p):
#P = np.dot(X,v)
#P = vecProd(X1, X2, v)
P = sparse_kronecker_multiplication_tools_python.x_gets_subset_of_A_kron_B_times_v(
v, X2, X1.T, colind, rowind)
z = (1. - Y * P)
z = np.where(z > 0, z, 0)
sv = np.nonzero(z)[0]
#map to rows and cols
rows = rowind[sv]
cols = colind[sv]
p_after = sparse_kronecker_multiplication_tools_python.x_gets_subset_of_A_kron_B_times_v(
p, X2, X1.T, cols, rows)
p_after = sparse_kronecker_multiplication_tools_python.x_gets_A_kron_B_times_sparse_v(
p_after, X1.T, X2, rows, cols)
p_after = p_after.reshape(X2.shape[1], X1.shape[1]).T.ravel()
return 2 * p_after + lamb * p
def predictWithDataMatricesAlt(self, X1pred, X2pred, row_inds = None, col_inds = None):
if row_inds == None:
P = np.dot(np.dot(X1pred, self.W), X2pred.T)
P = P.reshape(X1pred.shape[0] * X2pred.shape[0], 1, order = 'F')
else:
P = sparse_kronecker_multiplication_tools_python.x_gets_subset_of_A_kron_B_times_v(self.W.reshape((self.W.shape[0] * self.W.shape[1], 1), order = 'F'), X1pred, X2pred.T, np.array(row_inds, dtype=np.int32), np.array(col_inds, dtype=np.int32))
#P = X1pred * self.W * X2pred.T
return P
def func(v, X1, X2, Y, rowind, colind, lamb):
#REPLACE
#P = np.dot(X,v)
P = sparse_kronecker_multiplication_tools_python.x_gets_subset_of_A_kron_B_times_v(v, X2, X1.T, colind, rowind)
z = (1. - Y*P)
#print z
z = np.where(z>0, z, 0)
#return np.dot(z,z)
return 0.5*(np.dot(z,z)+lamb*np.dot(v,v))
def gradient(v, X1, X2, Y, rowind, colind, lamb):
#REPLACE
#P = np.dot(X,v)
#P = vecProd(X1, X2, v)
P = sparse_kronecker_multiplication_tools_python.x_gets_subset_of_A_kron_B_times_v(v, X2, X1.T, colind, rowind)
z = (1. - Y*P)
z = np.where(z>0, z, 0)
sv = np.nonzero(z)[0]
#map to rows and cols
rows = rowind[sv]
cols = colind[sv]
#A = -2*np.dot(X[sv].T, Y[sv])
A = - sparse_kronecker_multiplication_tools_python.x_gets_A_kron_B_times_sparse_v(Y[sv], X1.T, X2, rows, cols)
A = A.reshape(X2.shape[1], X1.shape[1]).T.ravel()
#B = 2 * np.dot(X[sv].T, np.dot(X[sv],v))
v_after = sparse_kronecker_multiplication_tools_python.x_gets_subset_of_A_kron_B_times_v(v, X2, X1.T, cols, rows)
v_after = sparse_kronecker_multiplication_tools_python.x_gets_A_kron_B_times_sparse_v(v_after, X1.T, X2, rows, cols)
B = v_after.reshape(X2.shape[1], X1.shape[1]).T.ravel()
#print "FOOOBAAR"
return A + B + lamb*v
def primal_rls_objective(w, X1, X2, Y, rowind, colind, lamb):
#primal form of the objective function for regularized least squares
#w: current primal solution
#X1: samples x features data matrix for domain 1
#X2: samples x features data matrix for domain 2
#rowind: row indices for training pairs
#colind: column indices for training pairs
#lamb: regularization parameter
P = sparse_kronecker_multiplication_tools_python.x_gets_subset_of_A_kron_B_times_v(w, X2, X1.T, colind, rowind)
z = (Y - P)
return 0.5*(np.dot(z,z)+lamb*np.dot(w,w))
def primal_svm_objective(w, X1, X2, Y, rowind, colind, lamb):
#primal form of the objective function for support vector machine
#w: current primal solution
#X1: samples x features data matrix for domain 1
#X2: samples x features data matrix for domain 2
#rowind: row indices for training pairs
#colind: column indices for training pairs
#lamb: regularization parameter
#P = np.dot(X,v)
P = sparse_kronecker_multiplication_tools_python.x_gets_subset_of_A_kron_B_times_v(w, X2, X1.T, colind, rowind)
z = (1. - Y*P)
z = np.where(z>0, z, 0)
return 0.5*(np.dot(z,z)+lamb*np.dot(w,w))
def gradient(v):
#REPLACE
#P = np.dot(X,v)
#P = vecProd(X1, X2, v)
P = sparse_kronecker_multiplication_tools_python.x_gets_subset_of_A_kron_B_times_v(
v, X2, X1.T, colind, rowind)
z = (1. - Y * P)
z = np.where(z > 0, z, 0)
sv = np.nonzero(z)[0]
#map to rows and cols
rows = rowind[sv]
cols = colind[sv]
#A = -2*np.dot(X[sv].T, Y[sv])
A = -2 * sparse_kronecker_multiplication_tools_python.x_gets_A_kron_B_times_sparse_v(
Y[sv], X1.T, X2, rows, cols)
A = A.reshape(X2.shape[1], X1.shape[1]).T.ravel()
#B = 2 * np.dot(X[sv].T, np.dot(X[sv],v))
v_after = sparse_kronecker_multiplication_tools_python.x_gets_subset_of_A_kron_B_times_v(
v, X2, X1.T, cols, rows)
v_after = 2 * sparse_kronecker_multiplication_tools_python.x_gets_A_kron_B_times_sparse_v(
v_after, X1.T, X2, rows, cols)
B = v_after.reshape(X2.shape[1], X1.shape[1]).T.ravel()
#print "FOOOBAAR"
return A + B + lamb * v
def predictWithDataMatricesAlt(self,
X1pred,
X2pred,
row_inds=None,
col_inds=None):
if row_inds == None:
P = np.dot(np.dot(X1pred, self.W), X2pred.T)
P = P.reshape(X1pred.shape[0] * X2pred.shape[0], 1, order='F')
else:
P = sparse_kronecker_multiplication_tools_python.x_gets_subset_of_A_kron_B_times_v(
self.W.reshape((self.W.shape[0] * self.W.shape[1], 1),
order='F'), X1pred, X2pred.T,
np.array(row_inds, dtype=np.int32),
np.array(col_inds, dtype=np.int32))
#P = X1pred * self.W * X2pred.T
return P
def dual_from_primal(self):
w = self.W.ravel()
X1 = self.resource_pool['xmatrix1']
X2 = self.resource_pool['xmatrix2']
rowind = np.array(self.label_row_inds, dtype = np.int32)
colind = np.array(self.label_col_inds, dtype = np.int32)
P = sparse_kronecker_multiplication_tools_python.x_gets_subset_of_A_kron_B_times_v(w, X2, X1.T, colind, rowind)
#choose support vectors here?
K1 = self.resource_pool['kmatrix1']
K2 = self.resource_pool['kmatrix2']
ddim = len(rowind)
def mv(v):
return sparse_kronecker_multiplication_tools_python.x_gets_C_times_M_kron_N_times_B_times_v(v, K2, K1, rowind, colind, rowind, colind)
def rv(v):
return sparse_kronecker_multiplication_tools_python.x_gets_C_times_M_kron_N_times_B_times_v(v, K2, K1, rowind, colind, rowind, colind)
K = LinearOperator((ddim, ddim), matvec=mv, rmatvec=rv, dtype=np.float64)
#A = lsmr(K, P, maxiter=100)[0]
A = cg(K, P, maxiter=100)[0]
return KernelPairwiseModel(A, rowind, colind)
def mv(v):
v_after = sparse_kronecker_multiplication_tools_python.x_gets_subset_of_A_kron_B_times_v(v, X1, X2.T, label_row_inds, label_col_inds)
v_after = sparse_kronecker_multiplication_tools_python.x_gets_A_kron_B_times_sparse_v(v_after, X1.T, X2, label_row_inds, label_col_inds) + regparam * v
return v_after
