def decompositionFromPool(self, rpool):
"""Builds decomposition representing the training data from resource pool.
Default implementation
builds and decomposes the kernel matrix itself (standard case), or the
empirical kernel map of the training data, if reduced set approximation is
used. Inheriting classes may also re-implement this by decomposing the feature
map of the data (e.g. linear kernel with low-dimensional data).
@param rpool: resource pool
@type rpool: dict
@return: svals, evecs, U, Z
@rtype: tuple of numpy matrices
"""
train_X = rpool['X']
kernel = rpool['kernel_obj']
if 'basis_vectors' in rpool:
basis_vectors = rpool['basis_vectors']
if not train_X.shape[1] == basis_vectors.shape[1]:
raise Exception("X and basis_vectors have different number of columns")
K_r = kernel.getKM(train_X).T
Krr = kernel.getKM(basis_vectors)
svals, evecs, U, Z = decomposeSubsetKM(K_r, Krr)
else:
K = kernel.getKM(train_X).T
svals, evecs = linalg.eig_psd(K)
U, Z = None, None
return svals, evecs, U, Z
def decompositionFromPool(self, rpool):
kernel = rpool['kernel_obj']
self.X = array_tools.as_2d_array(rpool['X'], True)
if 'basis_vectors' in rpool:
basis_vectors = array_tools.as_2d_array(rpool['basis_vectors'], True)
if not self.X.shape[1] == basis_vectors.shape[1]:
raise Exception("X and basis_vectors have different number of columns")
else:
basis_vectors = None
if "bias" in rpool:
self.bias = float(rpool["bias"])
else:
self.bias = 1.
if basis_vectors is not None or self.X.shape[1] > self.X.shape[0]:
#First possibility: subset of regressors has been invoked
if basis_vectors is not None:
K_r = kernel.getKM(self.X).T
Krr = kernel.getKM(basis_vectors)
svals, evecs, U, Z = decomposeSubsetKM(K_r, Krr)
#Second possibility: dual mode if more attributes than examples
else:
K = kernel.getKM(self.X).T
svals, evecs = linalg.eig_psd(K)
U, Z = None, None
#Third possibility, primal decomposition
else:
#Invoking getPrimalDataMatrix adds the bias feature
X = getPrimalDataMatrix(self.X, self.bias)
evecs, svals, U = linalg.svd_economy_sized(X)
U, Z = None, None
return svals, evecs, U, Z
def decompositionFromPool(self, rpool):
"""Builds decomposition representing the training data from resource pool.
Default implementation
builds and decomposes the kernel matrix itself (standard case), or the
empirical kernel map of the training data, if reduced set approximation is
used. Inheriting classes may also re-implement this by decomposing the feature
map of the data (e.g. linear kernel with low-dimensional data).
@param rpool: resource pool
@type rpool: dict
@return: svals, evecs, U, Z
@rtype: tuple of numpy matrices
"""
train_X = rpool['X']
kernel = rpool['kernel_obj']
if 'basis_vectors' in rpool:
basis_vectors = rpool['basis_vectors']
if not train_X.shape[1] == basis_vectors.shape[1]:
raise Exception(
"X and basis_vectors have different number of columns")
K_r = kernel.getKM(train_X).T
Krr = kernel.getKM(basis_vectors)
svals, evecs, U, Z = decomposeSubsetKM(K_r, Krr)
else:
K = kernel.getKM(train_X).T
svals, evecs = linalg.eig_psd(K)
U, Z = None, None
return svals, evecs, U, Z
def decompositionFromPool(self, rpool):
kernel = rpool['kernel_obj']
self.X = array_tools.as_2d_array(rpool['X'], True)
if 'basis_vectors' in rpool:
basis_vectors = array_tools.as_2d_array(rpool['basis_vectors'],
True)
if not self.X.shape[1] == basis_vectors.shape[1]:
raise Exception(
"X and basis_vectors have different number of columns")
else:
basis_vectors = None
if "bias" in rpool:
self.bias = float(rpool["bias"])
else:
self.bias = 1.
if basis_vectors is not None or self.X.shape[1] > self.X.shape[0]:
#First possibility: subset of regressors has been invoked
if basis_vectors is not None:
K_r = kernel.getKM(self.X).T
Krr = kernel.getKM(basis_vectors)
svals, evecs, U, Z = decomposeSubsetKM(K_r, Krr)
#Second possibility: dual mode if more attributes than examples
else:
K = kernel.getKM(self.X).T
svals, evecs = linalg.eig_psd(K)
U, Z = None, None
#Third possibility, primal decomposition
else:
#Invoking getPrimalDataMatrix adds the bias feature
X = getPrimalDataMatrix(self.X, self.bias)
evecs, svals, U = linalg.svd_economy_sized(X)
U, Z = None, None
return svals, evecs, U, Z
def decompositionFromPool(self, rpool):
K_train = rpool['kernel_matrix']
if 'basis_vectors' in rpool:
if not K_train.shape[1] == rpool["basis_vectors"].shape[1]:
raise Exception("When using basis vectors, both kernel matrices must contain equal number of columns")
svals, rsvecs, U, Z = decomposeSubsetKM(K_train.T, rpool['basis_vectors'])
else:
svals, rsvecs = linalg.eig_psd(K_train)
U, Z = None, None
return svals, rsvecs, U, Z
def decompositionFromPool(self, rpool):
K_train = rpool['kernel_matrix']
if 'basis_vectors' in rpool:
if not K_train.shape[1] == rpool["basis_vectors"].shape[1]:
raise Exception(
"When using basis vectors, both kernel matrices must contain equal number of columns"
)
svals, rsvecs, U, Z = decomposeSubsetKM(K_train.T,
rpool['basis_vectors'])
else:
svals, rsvecs = linalg.eig_psd(K_train)
U, Z = None, None
return svals, rsvecs, U, Z
def solve(self, regparam):
"""Re-trains RankRLS for the given regparam.
Parameters
----------
regparam : float, optional
regularization parameter, regparam > 0 (default=1.0)
Notes
-----
"""
size = self.svecs.shape[0]
if not hasattr(self, "multipleright"):
vals = np.concatenate([
np.ones((self.pairs.shape[0]), dtype=np.float64), -np.ones(
(self.pairs.shape[0]), dtype=np.float64)
])
row = np.concatenate([
np.arange(self.pairs.shape[0]),
np.arange(self.pairs.shape[0])
])
col = np.concatenate([self.pairs[:, 0], self.pairs[:, 1]])
coo = coo_matrix((vals, (row, col)),
shape=(self.pairs.shape[0], size))
self.L = (coo.T * coo) #.todense()
#Eigenvalues of the kernel matrix
evals = np.multiply(self.svals, self.svals)
#Temporary variables
ssvecs = np.multiply(self.svecs, self.svals)
#These are cached for later use in solve and computeHO functions
ssvecsTLssvecs = ssvecs.T * self.L * ssvecs
LRsvals, LRevecs = linalg.eig_psd(ssvecsTLssvecs)
LRsvals = np.mat(LRsvals)
LRevals = np.multiply(LRsvals, LRsvals)
LY = coo.T * np.mat(np.ones((self.pairs.shape[0], 1)))
self.multipleright = LRevecs.T * (ssvecs.T * LY)
self.multipleleft = ssvecs * LRevecs
self.LRevals = LRevals
self.LRevecs = LRevecs
self.regparam = regparam
#Compute the eigenvalues determined by the given regularization parameter
self.neweigvals = 1. / (self.LRevals + regparam)
self.A = self.svecs * np.multiply(1. / self.svals.T, (
self.LRevecs * np.multiply(self.neweigvals.T, self.multipleright)))
self.predictor = self.svdad.createModel(self)
def solve(self, regparam):
"""Re-trains RankRLS for the given regparam.
Parameters
----------
regparam : float, optional
regularization parameter, regparam > 0 (default=1.0)
Notes
-----
"""
size = self.svecs.shape[0]
if not hasattr(self, "multipleright"):
vals = np.concatenate(
[np.ones((self.pairs.shape[0]), dtype=np.float64), -np.ones((self.pairs.shape[0]), dtype=np.float64)]
)
row = np.concatenate([np.arange(self.pairs.shape[0]), np.arange(self.pairs.shape[0])])
col = np.concatenate([self.pairs[:, 0], self.pairs[:, 1]])
coo = coo_matrix((vals, (row, col)), shape=(self.pairs.shape[0], size))
self.L = coo.T * coo # .todense()
# Eigenvalues of the kernel matrix
evals = np.multiply(self.svals, self.svals)
# Temporary variables
ssvecs = np.multiply(self.svecs, self.svals)
# These are cached for later use in solve and computeHO functions
ssvecsTLssvecs = ssvecs.T * self.L * ssvecs
LRsvals, LRevecs = linalg.eig_psd(ssvecsTLssvecs)
LRsvals = np.mat(LRsvals)
LRevals = np.multiply(LRsvals, LRsvals)
LY = coo.T * np.mat(np.ones((self.pairs.shape[0], 1)))
self.multipleright = LRevecs.T * (ssvecs.T * LY)
self.multipleleft = ssvecs * LRevecs
self.LRevals = LRevals
self.LRevecs = LRevecs
self.regparam = regparam
# Compute the eigenvalues determined by the given regularization parameter
self.neweigvals = 1.0 / (self.LRevals + regparam)
self.A = self.svecs * np.multiply(
1.0 / self.svals.T, (self.LRevecs * np.multiply(self.neweigvals.T, self.multipleright))
)
self.predictor = self.svdad.createModel(self)
def solve(self, regparam=1.0):
"""Trains the learning algorithm, using the given regularization parameter.
Parameters
----------
regparam : float (regparam > 0)
regularization parameter
Notes
-----
Computational complexity of re-training:
m = n_samples, d = n_features, l = n_labels, b = n_bvectors
O(lm^2): basic case
O(lmd): Linear Kernel, d < m
O(lmb): Sparse approximation with basis vectors
"""
if not hasattr(self, "D"):
qidlist = self.qids
objcount = max(qidlist) + 1
labelcounts = np.mat(np.zeros((1, objcount)))
Pvals = np.ones(self.size)
for i in range(self.size):
qid = qidlist[i]
labelcounts[0, qid] = labelcounts[0, qid] + 1
D = np.mat(np.ones((1, self.size), dtype=np.float64))
#The centering matrix way (HO computations should be modified accordingly too)
for i in range(self.size):
qid = qidlist[i]
Pvals[i] = 1. / np.sqrt(labelcounts[0, qid])
#The old Laplacian matrix way
#for i in range(self.size):
# qid = qidlist[i]
# D[0, i] = labelcounts[0, qid]
P = scipy.sparse.coo_matrix(
(Pvals, (np.arange(0, self.size), qidlist)),
shape=(self.size, objcount))
P_csc = P.tocsc()
P_csr = P.tocsr()
#Eigenvalues of the kernel matrix
#evals = np.multiply(self.svals, self.svals)
#Temporary variables
ssvecs = np.multiply(self.svecs, self.svals)
#These are cached for later use in solve and holdout functions
ssvecsTLssvecs = (np.multiply(ssvecs.T, D) -
(ssvecs.T * P_csc) * P_csr.T) * ssvecs
LRsvals, LRevecs = linalg.eig_psd(ssvecsTLssvecs)
LRsvals = np.mat(LRsvals)
LRevals = np.multiply(LRsvals, LRsvals)
LY = np.multiply(D.T, self.Y) - P_csr * (P_csc.T * self.Y)
self.multipleright = LRevecs.T * (ssvecs.T * LY)
self.multipleleft = ssvecs * LRevecs
self.LRevals = LRevals
self.LRevecs = LRevecs
self.D = D
self.regparam = regparam
#Compute the eigenvalues determined by the given regularization parameter
self.neweigvals = 1. / (self.LRevals + regparam)
self.A = self.svecs * np.multiply(1. / self.svals.T, (
self.LRevecs * np.multiply(self.neweigvals.T, self.multipleright)))
self.predictor = self.svdad.createModel(self)
def solve(self, regparam1, regparam2):
"""Re-trains TwoStepRLS for the given regparams
Parameters
----------
regparam1: float
regularization parameter 1, regparam1 > 0
regparam2: float
regularization parameter 2, regparam2 > 0
Notes
-----
Computational complexity of re-training:
m = n_samples1, n = n_samples2, d = n_features1, e = n_features2
O(ed^2 + de^2) Linear version (assumption: d < m, e < n)
O(m^3 + n^3) Kernel version
"""
self.regparam1 = regparam1
self.regparam2 = regparam2
if self.kernelmode:
K1, K2 = self.K1, self.K2
Y = self.Y.reshape((K1.shape[0], K2.shape[0]), order='F')
if not self.trained:
self.trained = True
evals1, V = linalg.eig_psd(K1)
evals1 = np.mat(evals1).T
evals1 = np.multiply(evals1, evals1)
V = np.mat(V)
self.evals1 = evals1
self.V = V
evals2, U = linalg.eig_psd(K2)
evals2 = np.mat(evals2).T
evals2 = np.multiply(evals2, evals2)
U = np.mat(U)
self.evals2 = evals2
self.U = U
self.VTYU = V.T * self.Y * U
self.newevals1 = 1. / (self.evals1 + regparam1)
self.newevals2 = 1. / (self.evals2 + regparam2)
newevals = self.newevals1 * self.newevals2.T
self.A = np.multiply(self.VTYU, newevals)
self.A = self.V * self.A * self.U.T
self.A = np.array(self.A)
label_row_inds, label_col_inds = np.unravel_index(np.arange(K1.shape[0] * K2.shape[0]), (K1.shape[0], K2.shape[0]))
label_row_inds = np.array(label_row_inds, dtype = np.int32)
label_col_inds = np.array(label_col_inds, dtype = np.int32)
self.predictor = KernelPairwisePredictor(self.A.ravel(), label_row_inds, label_col_inds)
else:
X1, X2 = self.X1, self.X2
Y = self.Y.reshape((X1.shape[0], X2.shape[0]), order='F')
if not self.trained:
self.trained = True
V, svals1, rsvecs1 = linalg.svd_economy_sized(X1)
svals1 = np.mat(svals1)
self.svals1 = svals1.T
self.evals1 = np.multiply(self.svals1, self.svals1)
self.V = V
self.rsvecs1 = np.mat(rsvecs1)
if X1.shape == X2.shape and (X1 == X2).all():
svals2, U, rsvecs2 = svals1, V, rsvecs1
else:
U, svals2, rsvecs2 = linalg.svd_economy_sized(X2)
svals2 = np.mat(svals2)
self.svals2 = svals2.T
self.evals2 = np.multiply(self.svals2, self.svals2)
self.U = U
self.rsvecs2 = np.mat(rsvecs2)
self.VTYU = V.T * Y * U
self.newevals1 = 1. / (self.evals1 + regparam1)
self.newevals2 = 1. / (self.evals2 + regparam2)
newevals = np.multiply(self.svals1, self.newevals1) * np.multiply(self.svals2, self.newevals2).T
self.W = np.multiply(self.VTYU, newevals)
self.W = self.rsvecs1.T * self.W * self.rsvecs2
#self.predictor = LinearPairwisePredictor(self.W)
self.predictor = LinearPairwisePredictor(np.array(self.W))
def solve(self, regparam=1.0):
"""Trains the learning algorithm, using the given regularization parameter.
Parameters
----------
regparam : float (regparam > 0)
regularization parameter
Notes
-----
Computational complexity of re-training:
m = n_samples, d = n_features, l = n_labels, b = n_bvectors
O(lm^2): basic case
O(lmd): Linear Kernel, d < m
O(lmb): Sparse approximation with basis vectors
"""
if not hasattr(self, "D"):
qidlist = self.qids
objcount = max(qidlist) + 1
labelcounts = np.mat(np.zeros((1, objcount)))
Pvals = np.ones(self.size)
for i in range(self.size):
qid = qidlist[i]
labelcounts[0, qid] = labelcounts[0, qid] + 1
D = np.mat(np.ones((1, self.size), dtype=np.float64))
#The centering matrix way (HO computations should be modified accordingly too)
for i in range(self.size):
qid = qidlist[i]
Pvals[i] = 1. / np.sqrt(labelcounts[0, qid])
#The old Laplacian matrix way
#for i in range(self.size):
# qid = qidlist[i]
# D[0, i] = labelcounts[0, qid]
P = scipy.sparse.coo_matrix((Pvals, (np.arange(0, self.size), qidlist)), shape=(self.size,objcount))
P_csc = P.tocsc()
P_csr = P.tocsr()
#Eigenvalues of the kernel matrix
#evals = np.multiply(self.svals, self.svals)
#Temporary variables
ssvecs = np.multiply(self.svecs, self.svals)
#These are cached for later use in solve and holdout functions
ssvecsTLssvecs = (np.multiply(ssvecs.T, D) - (ssvecs.T * P_csc) * P_csr.T) * ssvecs
LRsvals, LRevecs = linalg.eig_psd(ssvecsTLssvecs)
LRsvals = np.mat(LRsvals)
LRevals = np.multiply(LRsvals, LRsvals)
LY = np.multiply(D.T, self.Y) - P_csr * (P_csc.T * self.Y)
self.multipleright = LRevecs.T * (ssvecs.T * LY)
self.multipleleft = ssvecs * LRevecs
self.LRevals = LRevals
self.LRevecs = LRevecs
self.D = D
self.regparam = regparam
#Compute the eigenvalues determined by the given regularization parameter
self.neweigvals = 1. / (self.LRevals + regparam)
self.A = self.svecs * np.multiply(1. / self.svals.T, (self.LRevecs * np.multiply(self.neweigvals.T, self.multipleright)))
self.predictor = self.svdad.createModel(self)
def solve(self, regparam1, regparam2):
"""Re-trains TwoStepRLS for the given regparams
Parameters
----------
regparam1: float
regularization parameter 1, regparam1 > 0
regparam2: float
regularization parameter 2, regparam2 > 0
Notes
-----
Computational complexity of re-training:
m = n_samples1, n = n_samples2, d = n_features1, e = n_features2
O(ed^2 + de^2) Linear version (assumption: d < m, e < n)
O(m^3 + n^3) Kernel version
"""
self.regparam1 = regparam1
self.regparam2 = regparam2
if self.kernelmode:
K1, K2 = self.K1, self.K2
Y = self.Y.reshape((K1.shape[0], K2.shape[0]), order='F')
if not self.trained:
self.trained = True
evals1, V = linalg.eig_psd(K1)
evals1 = np.mat(evals1).T
evals1 = np.multiply(evals1, evals1)
V = np.mat(V)
self.evals1 = evals1
self.V = V
evals2, U = linalg.eig_psd(K2)
evals2 = np.mat(evals2).T
evals2 = np.multiply(evals2, evals2)
U = np.mat(U)
self.evals2 = evals2
self.U = U
self.VTYU = V.T * self.Y * U
self.newevals1 = 1. / (self.evals1 + regparam1)
self.newevals2 = 1. / (self.evals2 + regparam2)
newevals = self.newevals1 * self.newevals2.T
self.A = np.multiply(self.VTYU, newevals)
self.A = self.V * self.A * self.U.T
self.A = np.array(self.A)
label_row_inds, label_col_inds = np.unravel_index(
np.arange(K1.shape[0] * K2.shape[0]),
(K1.shape[0], K2.shape[0]))
label_row_inds = np.array(label_row_inds, dtype=np.int32)
label_col_inds = np.array(label_col_inds, dtype=np.int32)
self.predictor = KernelPairwisePredictor(self.A.ravel(),
label_row_inds,
label_col_inds)
else:
X1, X2 = self.X1, self.X2
Y = self.Y.reshape((X1.shape[0], X2.shape[0]), order='F')
if not self.trained:
self.trained = True
V, svals1, rsvecs1 = linalg.svd_economy_sized(X1)
svals1 = np.mat(svals1)
self.svals1 = svals1.T
self.evals1 = np.multiply(self.svals1, self.svals1)
self.V = V
self.rsvecs1 = np.mat(rsvecs1)
if X1.shape == X2.shape and (X1 == X2).all():
svals2, U, rsvecs2 = svals1, V, rsvecs1
else:
U, svals2, rsvecs2 = linalg.svd_economy_sized(X2)
svals2 = np.mat(svals2)
self.svals2 = svals2.T
self.evals2 = np.multiply(self.svals2, self.svals2)
self.U = U
self.rsvecs2 = np.mat(rsvecs2)
self.VTYU = V.T * Y * U
self.newevals1 = 1. / (self.evals1 + regparam1)
self.newevals2 = 1. / (self.evals2 + regparam2)
newevals = np.multiply(self.svals1, self.newevals1) * np.multiply(
self.svals2, self.newevals2).T
self.W = np.multiply(self.VTYU, newevals)
self.W = self.rsvecs1.T * self.W * self.rsvecs2
#self.predictor = LinearPairwisePredictor(self.W)
self.predictor = LinearPairwisePredictor(np.array(self.W))
