def _get_gramians(self, X):
"""
Parameters
----------
X: array, (n_dims, n_samples)
Returns
-------
G: array, (n_class, n_subdims, n_subdims)
gramian matricies of references of each class
"""
# _X, (n_dims, n_samples)
# K, (n_samples, n_samples)
K = rbf_kernel(X, X, self.sigma)
# in_coeff, (n_samles, test_n_subdims)
in_coeff, _ = dual_vectors(K, self.test_n_subdims)
gramians = []
for i in range(self.n_data):
# ref_X, (n_dims, n_samples_ref_X)
# ref_coeff, (n_samples_ref_X, n_subdims)
ref_X, ref_coeff = self.dic[i]
# _K, (n_samples_ref_X, n_samples)
_K = rbf_kernel(ref_X, X, self.sigma)
# S, (n_subdims, test_n_subdims)
S = ref_coeff.T.dot(_K.dot(in_coeff))
gramians.append(S)
return np.array(gramians)
def fast_predict_proba(self, X):
"""
Predict class probabilities
Parameters:
-----------
X: list of 2d-arrays, (n_vector_sets, n_samples, n_dims)
List of input vector sets.
Returns:
--------
pred: array, (n_vector_sets)
Prediction array
"""
n_input = len(X)
n_ref = len(self.dic)
# preprocessing data matricies
X = self._prepare_X(X)
# manage reference informations
ref_Xs, ref_coeffs = [], []
for ref_X, ref_coeff in self.dic:
ref_Xs.append(ref_X)
ref_coeffs.append(ref_coeff)
ref_mappings = np.array([
i for i in range(len(ref_Xs))
for _ in range(ref_coeffs[i].shape[1])
])
ref_Xs = np.hstack(ref_Xs)
ref_coeffs = block_diag(*ref_coeffs)
in_coeffs = []
for _X in X:
K = rbf_kernel(_X, _X, self.sigma)
in_coeff, _ = dual_vectors(K, self.n_subdims)
in_coeffs.append(in_coeff)
in_mappings = np.array(
[i for i in range(n_input) for _ in range(in_coeffs[i].shape[1])])
in_Xs = np.hstack(X)
in_coeffs = block_diag(*in_coeffs)
K = rbf_kernel(in_Xs, ref_Xs)
del ref_Xs, in_Xs
S = in_coeffs.T.dot(K).dot(ref_coeffs)
del in_coeffs, ref_coeffs, K
pred = np.zeros((n_input, n_ref), dtype=np.float64)
for i in range(n_input):
for j in range(n_ref):
pred[i, j] = mean_square_singular_values(
S[in_mappings == i][:, ref_mappings == j])
del S, X
return np.array(pred)
def _fit(self, X, y):
"""
Parameters
----------
X: list of 2d-arrays, (n_classes, n_dims, n_samples)
y: array, (n_classes)
"""
coeff = []
for _X in X:
K = rbf_kernel(_X, _X, self.sigma)
_coeff, _ = dual_vectors(K, self.n_subdims)
coeff.append(_coeff)
self.dic = list(zip(X, coeff))
def _fit(self, X, y):
"""
Parameters
----------
X: list of 2d-arrays, (n_classes, n_dims, n_samples)
y: array, (n_classes)
"""
# mapings, (n_classes * n_samples)
mappings = np.array([
i for i, _X in enumerate(X)
for _ in range(_X.shape[1])
])
# stack_X, (n_dims, n_classes * n_samples)
stack_X = np.hstack(X)
# K, (n_classes * n_samples, n_classes * n_samples)
K = rbf_kernel(stack_X, stack_X, self.sigma)
coeff = []
for i, _ in enumerate(X):
p = (mappings == i)
# clipping K(X[y==c], X[y==c])
# _K, (n_samples_i, n_samples_i)
_K = K[p][:, p]
_coeff, _ = dual_vectors(_K, self.n_subdims)
coeff.append(_coeff)
# coeff, (n_classes * n_samples, n_classes * n_subdims)
coeff = block_diag(*coeff)
# gramian, (n_classes * n_subdims, n_classes * n_subdims)
gramian = coeff.T @ K @ coeff
# n_gds_dims
if 0.0 < self.n_gds_dims <= 1.0:
n_gds_dims = int(gramian.shape[0] * self.n_gds_dims)
else:
n_gds_dims = self.n_gds_dims
# coefficients of `bases in feature space` to get GDS bases in feature space
# gds_coeff, (n_classes * n_subdims, n_gds_dims)
gds_coeff, _ = dual_vectors(gramian, n_gds_dims, higher=False)
# coefficients of `given data X in feature space` to get GDS bases in features space
# gds_coeff, (n_classes * n_samples, n_gds_dims)
gds_coeff = np.dot(coeff, gds_coeff)
# X_gds = (GDS bases in feature space).T @ (X_stacked in feature space)
# projection coefficients of GDS in feature space.
# this vectors have finite dimension, such as n_gds_dims,
# and these vectors can treat as linear feature.
# X_gds, (n_gds_dims, n_classes * n_samples)
X_gds = np.dot(gds_coeff.T, K)
# X_gds, (n_classes, n_gds_dims, n_samples)
X_gds = [X_gds[:, mappings == i] for i, _ in enumerate(X)]
# dic, (n_classes, n_dims, n_subdims)
dic = [subspace_bases(_X, self.n_subdims) for _X in X_gds]
dic = np.array(dic)
self.train_stack_X = stack_X
self.mappings = mappings
self.gds_coeff = gds_coeff
self.dic = dic
def fast_predict_proba(self, X):
"""
Predict class probabilities
Parameters:
-----------
X: list of 2d-arrays, (n_vector_sets, n_samples, n_dims)
List of input vector sets.
Returns:
--------
pred: array, (n_vector_sets)
Prediction array
"""
n_input = len(X)
n_ref = len(self.dic)
# preprocessing data matricies
X = self._prepare_X(X)
# manage reference informations
ref_Xs, ref_coeffs = [], []
for ref_X, ref_coeff in self.dic:
ref_Xs.append(ref_X)
ref_coeffs.append(ref_coeff)
ref_mappings = np.array([
i for i in range(len(ref_Xs))
for _ in range(ref_coeffs[i].shape[1])
])
ref_Xs = np.hstack(ref_Xs)
ref_coeffs = block_diag(*ref_coeffs)
in_coeffs = []
for _X in X:
K = rbf_kernel(_X, _X, self.sigma)
in_coeff, _ = dual_vectors(K, self.test_n_subdims)
in_coeffs.append(in_coeff)
in_mappings = np.array(
[i for i in range(n_input) for _ in range(in_coeffs[i].shape[1])])
in_Xs = np.hstack(X)
in_coeffs = block_diag(*in_coeffs)
K = rbf_kernel(in_Xs, ref_Xs, self.sigma)
del ref_Xs, in_Xs
S = in_coeffs.T.dot(K).dot(ref_coeffs)
del in_coeffs, ref_coeffs, K
# Split matrix into (n_input x n_ref) blocks
in_split = np.where(np.diff(np.pad(in_mappings, (1, 0),
'constant')))[0]
ref_split = np.where(np.diff(np.pad(ref_mappings, (1, 0),
'constant')))[0]
S = [np.hsplit(_S, ref_split) for _S in np.vsplit(S, in_split)]
vmssv = np.vectorize(lambda i, j: mean_square_singular_values(S[i][j]))
pred = vmssv(*np.meshgrid(np.arange(n_input), np.arange(n_ref))).T
del S, X
return np.array(pred)