def predict(self, X):
"""
Make predictions on the new samplets in X.
For an one-class model, +1 or -1 is returned.
Parameters
----------
X : {array-like, sparse matrix}, shape (n_samples, n_features)
For kernel="precomputed", the expected shape of X is
[n_samples_test, n_samples_train]
Returns
-------
y_pred : array, shape (n_samples,)
Class labels for samples in X.
"""
if not hasattr(self, '_km'):
raise NotFittedError("Can't predict. Not fitted yet. Run .fit() first!")
test_X = check_array(X)
# this is a fresh new KM
self._km = KernelMatrix(self.k_func, name='test_km',
normalized=self.normalized)
# sample_one must be test data to get the right shape for sklearn X
self._km.attach_to(sample_one=test_X, sample_two=self._train_X)
predicted_y = self._estimator.predict(self._km.full)
return np.asarray(predicted_y, dtype=self._train_y.dtype)
def test_size_property_mismatch():
ks = KernelSet(num_samples=sample_data.shape[0] + 1)
lin = KernelMatrix(LinearKernel(skip_input_checks=True))
lin.attach_to(sample_data)
with raises(KMSetAdditionError):
ks.append(lin)
def _test_func_is_valid_kernel(kernel, sample_dim, num_samples):
"""A func is a valid kernel if the kernel matrix generated by it is PSD.
Not including this in tests for all kernels to allow for non-PSD kernels in the future
"""
KM = KernelMatrix(kernel, name='TestKM')
KM.attach_to(gen_random_sample(num_samples, sample_dim))
is_psd = is_positive_semidefinite(KM.full, verbose=True)
if not is_psd:
raise ValueError('{} is not PSD'.format(str(KM)))
def fit(self, X, y, sample_weight=None):
"""Fit the chosen Estimator based on the user-defined kernel.
Parameters
----------
X : {array-like, sparse matrix}, shape (n_samples, n_features)
Training vectors, where n_samples is the number of samples
and n_features is the number of features.
y : array-like, shape (n_samples,)
Target values (class labels in classification, real numbers in
regression)
sample_weight : array-like, shape (n_samples,)
Per-sample weights. Rescale C per sample. Higher weights
force the classifier to put more emphasis on these points.
Returns
-------
self : object
Notes
------
If X and y are not C-ordered and contiguous arrays of np.float64 and
X is not a scipy.sparse.csr_matrix, X and/or y may be copied.
If X is a dense array, then the other methods will not support sparse
matrices as input.
"""
if is_regressor(self):
self._train_X, self._train_y = check_X_y(X, y, y_numeric=True)
self._train_y = self._train_y.astype(np.float_)
else:
self._train_X, self._train_y = check_X_y(X, y)
self._km = KernelMatrix(self.k_func,
name='train_km',
normalized=self.normalized)
self._km.attach_to(self._train_X)
self._estimator, self.param_grid = get_estimator(self.learner_id)
self._estimator.fit(X=self._km.full,
y=self._train_y,
sample_weight=sample_weight)
if is_classifier(self):
self.classes_ = self._estimator.classes_
return self
def test_KernelMatrix_design():
with raises(TypeError):
km = KernelMatrix(kernel=simple_callable)
with raises(TypeError):
km = KernelMatrix(kernel=LinearKernel, normalized='True')
assert len(km_lin) == num_samples**2
colon_access = km_lin[:, :]
if colon_access.size != km_lin.size:
raise ValueError('error in getitem implementation when using [:, :]')
_ = km_lin[1, :]
_ = km_lin[:, 1]
for invalid_index in (-1, np.Inf, np.NaN):
with raises(KMAccessError):
_ = km_lin[:, invalid_index]
def add_parametrized_kernels(self, kernel_func, param, values):
"""
Adds a list of kernels parametrized by various values for a given param
Parameters
----------
kernel_func : BaseKernelFunction
Kernel function to be added (not an instance, but callable class)
param : str
Name of the parameter to the above kernel function
values : Iterable
List of parameter values. One kernel will be added for each value
"""
if (not isinstance(kernel_func, type)) or \
(not issubclass(kernel_func, BaseKernelFunction)):
raise KernelMethodsException(
'Input {} is not a valid kernel func!'
' Must be derived from BaseKernelFunction'
''.format(kernel_func))
if values is None:
# warn('No values provided for {}. Doing nothing!'.format(param))
return
if not is_iterable_but_not_str(values, min_length=1):
raise ValueError(
'values must be an iterable set of param values (n>=1)')
for val in values:
try:
param_dict = {
param: val,
'skip_input_checks': self._skip_input_checks
}
self.append(
KernelMatrix(kernel_func(**param_dict),
normalized=self._norm_kernels))
except:
warn(
'Unable to add {} to the bucket for {}={}. Skipping it.'
''.format(kernel_func, param, val), KernelMethodsWarning)
def test_alignment_centered():
km1 = KernelMatrix(kernel=LinearKernel())
km1.attach_to(gen_random_sample(num_samples, sample_dim))
km2 = KernelMatrix(kernel=LinearKernel())
km2.attach_to(gen_random_sample(num_samples, sample_dim))
km3_bad_size = KernelMatrix(kernel=LinearKernel())
km3_bad_size.attach_to(gen_random_sample(num_samples + 2, sample_dim))
with raises(ValueError):
alignment_centered(km1.full, km3_bad_size.full)
# bad type : must be ndarray
with raises(TypeError):
alignment_centered(km1, km2.full)
# bad type : must be ndarray
with raises(TypeError):
alignment_centered(km1.full, km2)
for flag in (True, False):
_ = alignment_centered(km1.full, km2.full, centered_already=flag)
with raises(ValueError):
_ = alignment_centered(np.zeros((10, 10)),
randn(10, 10),
value_if_zero_division='raise')
return_val_requested = 'random_set_value'
with warns(UserWarning):
ret_value = alignment_centered(
randn(10, 10),
np.zeros((10, 10)),
value_if_zero_division=return_val_requested)
if ret_value != return_val_requested:
raise ValueError('Not returning the value requested in case of error!')
from kernelmethods.base import KMSetAdditionError, KernelMatrix, KernelSet, \
BaseKernelFunction
from kernelmethods.numeric_kernels import GaussianKernel, LinearKernel, PolyKernel
from kernelmethods.sampling import make_kernel_bucket
num_samples = 50 # 9
sample_dim = 3 # 2
target_label_set = [1, 2]
sample_data = np.random.rand(num_samples, sample_dim)
target_labels = np.random.choice(target_label_set, (num_samples, 1))
IdealKM = target_labels.dot(target_labels.T)
rbf = KernelMatrix(GaussianKernel(sigma=10, skip_input_checks=True))
lin = KernelMatrix(LinearKernel(skip_input_checks=True))
poly = KernelMatrix(PolyKernel(degree=2, skip_input_checks=True))
# lin.attach_to(sample_data)
# rbf.attach_to(sample_data)
# poly.attach_to(sample_data)
kset = KernelSet([lin, poly, rbf])
print(kset)
def test_creation():
try:
ks = KernelSet()
class BaseKernelMachine(BaseEstimator):
"""Generic class to return a drop-in sklearn estimator.
Parameters
----------
k_func : KernelFunction
The kernel function the kernel machine bases itself on
learner_id : str
Identifier for the estimator to be built based on the kernel function.
Options: ``SVC`` and ``SVR``.
Default: ``SVC`` (classifier version of SVM)
normalized : flag
Flag to indicate whether to keep the kernel matrix normalized
Default: False
"""
def __init__(self,
k_func=GaussianKernel(),
learner_id='SVC',
normalized=False):
"""
Constructor for the KernelMachine class.
Parameters
----------
k_func : KernelFunction
The kernel function the kernel machine bases itself on
learner_id : str
Identifier for the estimator to be built based on the kernel function.
Options: ``SVC`` and ``SVR``.
Default: ``SVC`` (classifier version of SVM)
normalized : flag
Flag to indicate whether to keep the kernel matrix normalized.
Default: False
"""
self.k_func = k_func
self.learner_id = learner_id
self.normalized = normalized
self._estimator, self.param_grid = get_estimator(self.learner_id)
def fit(self, X, y, sample_weight=None):
"""Fit the chosen Estimator based on the user-defined kernel.
Parameters
----------
X : {array-like, sparse matrix}, shape (n_samples, n_features)
Training vectors, where n_samples is the number of samples
and n_features is the number of features.
For kernel="precomputed", the expected shape of X is
(n_samples, n_samples).
y : array-like, shape (n_samples,)
Target values (class labels in classification, real numbers in
regression)
sample_weight : array-like, shape (n_samples,)
Per-sample weights. Rescale C per sample. Higher weights
force the classifier to put more emphasis on these points.
Returns
-------
self : object
Notes
------
If X and y are not C-ordered and contiguous arrays of np.float64 and
X is not a scipy.sparse.csr_matrix, X and/or y may be copied.
If X is a dense array, then the other methods will not support sparse
matrices as input.
"""
if is_regressor(self):
self._train_X, self._train_y = check_X_y(X, y, y_numeric=True)
self._train_y = self._train_y.astype(np.float_)
else:
self._train_X, self._train_y = check_X_y(X, y)
self._km = KernelMatrix(self.k_func, name='train_km',
normalized=self.normalized)
self._km.attach_to(self._train_X)
self._estimator.fit(X=self._km.full, y=self._train_y,
sample_weight=sample_weight)
if is_classifier(self):
self.classes_ = self._estimator.classes_
return self
def predict(self, X):
"""
Make predictions on the new samplets in X.
For an one-class model, +1 or -1 is returned.
Parameters
----------
X : {array-like, sparse matrix}, shape (n_samples, n_features)
For kernel="precomputed", the expected shape of X is
[n_samples_test, n_samples_train]
Returns
-------
y_pred : array, shape (n_samples,)
Class labels for samples in X.
"""
if not hasattr(self, '_km'):
raise NotFittedError("Can't predict. Not fitted yet. Run .fit() first!")
test_X = check_array(X)
# this is a fresh new KM
self._km = KernelMatrix(self.k_func, name='test_km',
normalized=self.normalized)
# sample_one must be test data to get the right shape for sklearn X
self._km.attach_to(sample_one=test_X, sample_two=self._train_X)
predicted_y = self._estimator.predict(self._km.full)
return np.asarray(predicted_y, dtype=self._train_y.dtype)
def get_params(self, deep=True):
"""returns all the relevant parameters for this estimator!"""
return {'k_func' : self.k_func,
'normalized': self.normalized,
'learner_id': self.learner_id}
def set_params(self, **parameters):
"""Param setter"""
for parameter, value in parameters.items():
if parameter in ('k_func', 'learner_id', 'normalized'):
setattr(self, parameter, value)
return self
def test_normalize():
km = KernelMatrix(kernel=LinearKernel())
km.attach_to(gen_random_sample(num_samples, sample_dim))
km.normalize()
def test_centering():
km = KernelMatrix(kernel=LinearKernel())
km.attach_to(gen_random_sample(num_samples, sample_dim))
km.center()
def gen_random_array(dim):
"""To better control precision and type of floats"""
# TODO input sparse arrays for test
return np.random.rand(dim)
def gen_random_sample(num_samples, sample_dim):
"""To better control precision and type of floats"""
# TODO input sparse arrays for test
return np.random.rand(num_samples, sample_dim)
km_lin = KernelMatrix(kernel=LinearKernel())
km_lin.attach_to(gen_random_sample(num_samples, sample_dim))
def simple_callable(x, y):
return np.dot(x, y)
def test_kernel_from_callable():
kf = KernelFromCallable(simple_callable)
if not isinstance(kf, BaseKernelFunction):
raise TypeError('Error in implementation of KernelFromCallable')
_test_for_all_kernels(kf, 5)
def __init__(
self,
poly_degree_values=cfg.default_degree_values_poly_kernel,
rbf_sigma_values=cfg.default_sigma_values_gaussian_kernel,
laplace_gamma_values=cfg.default_gamma_values_laplacian_kernel,
sigmoid_gamma_values=cfg.default_gamma_values_sigmoid_kernel,
sigmoid_offset_values=cfg.default_offset_values_sigmoid_kernel,
name='KernelBucket',
normalize_kernels=True,
skip_input_checks=False,
):
"""
Constructor.
Parameters
----------
poly_degree_values : Iterable
List of values for the degree parameter of the PolyKernel. One
KernelMatrix will be added to the bucket for each value.
rbf_sigma_values : Iterable
List of values for the sigma parameter of the GaussianKernel. One
KernelMatrix will be added to the bucket for each value.
laplace_gamma_values : Iterable
List of values for the gamma parameter of the LaplacianKernel. One
KernelMatrix will be added to the bucket for each value.
sigmoid_gamma_values : Iterable
List of values for the gamma parameter of the SigmoidKernel. One
KernelMatrix will be added to the bucket for each value.
sigmoid_offset_values : Iterable
List of values for the offset parameter of the SigmoidKernel. One
KernelMatrix will be added to the bucket for each value.
name : str
String to identify the purpose or type of the bucket of kernels.
Also helps easily distinguishing it from other buckets.
normalize_kernels : bool
Flag to indicate whether the kernel matrices need to be normalized
skip_input_checks : bool
Flag to indicate whether checks on input data (type, format etc) can
be skipped. This helps save a tiny bit of runtime for expert uses when
data types and formats are managed thoroughly in numpy. Default:
False. Disable this only when you know exactly what you're doing!
"""
if isinstance(normalize_kernels, bool):
self._norm_kernels = normalize_kernels
else:
raise TypeError('normalize_kernels must be bool')
if isinstance(skip_input_checks, bool):
self._skip_input_checks = skip_input_checks
else:
raise TypeError('skip_input_checks must be bool')
# start with the addition of kernel matrix for linear kernel
init_kset = [
KernelMatrix(LinearKernel(), normalized=self._norm_kernels),
]
super().__init__(km_list=init_kset, name=name)
# not attached to a sample yet
self._num_samples = None
self.add_parametrized_kernels(PolyKernel, 'degree', poly_degree_values)
self.add_parametrized_kernels(GaussianKernel, 'sigma',
rbf_sigma_values)
self.add_parametrized_kernels(LaplacianKernel, 'gamma',
laplace_gamma_values)
self.add_parametrized_kernels(SigmoidKernel, 'gamma',
sigmoid_gamma_values)
self.add_parametrized_kernels(SigmoidKernel, 'offset',
sigmoid_offset_values)
