""" One class variants of BRMM based on separation from zero """
# the output is a prediction vector
import logging
import warnings
import numpy
from pySPACE.missions.nodes.decorators import NoOptimizationParameter, BooleanParameter, NormalParameter, \
ChoiceParameter, LogUniformParameter
from pySPACE.resources.data_types.prediction_vector import PredictionVector
from pySPACE.missions.nodes.classification.one_class \
import OneClassClassifierBase
from pySPACE.missions.nodes.classification.svm_variants.brmm import RMM2Node, RmmPerceptronNode
@NoOptimizationParameter("offset_factor")
@BooleanParameter("outer_boundary")
[docs]class OcRmmNode(RMM2Node, OneClassClassifierBase):
""" Take zero as opposite class
**References**
========= ==================================================
main
========= ==================================================
author Krell, M. M. and Woehrle, H.
title `New one-class classifiers based on the origin separation approach <http://dx.doi.org/10.1016/j.patrec.2014.11.008>`_
journal Pattern Recognition Letters
publisher Elsevier
doi 10.1016/j.patrec.2014.11.008
year 2015
========= ==================================================
**Exemplary Call**
.. code-block:: yaml
- node : OcRmmNode
parameters :
complexity : 1.0
class_labels : ['Target']
range : 4
:input: FeatureVector
:output: PredictionVector
:Author: Mario Michael Krell
:Created: 2013/08/16
"""
[docs] def __init__(self, outer_boundary=False, **kwargs):
if "offset_factor" in kwargs:
kwargs.pop("offset_factor")
RMM2Node.__init__(self, offset_factor=0, **kwargs)
self.set_permanent_attributes(
b=1, weight=[1, 1], one_class=True, outer_boundary=outer_boundary)
[docs] def append_weights_and_class_factors(self, label):
""" Only label zero is expected and label factor one is used """
if not label == 0:
self._log("Unexpected label (%s) occurred!" % str(label),
level=logging.ERROR)
self.bi.append(-1)
if self.linear_weighting:
self.ci.append(self.complexity*self.weight[0]*self.num_samples)
else:
self.ci.append(self.complexity*self.weight[0])
[docs] def train(self, data, label):
""" Forward to one class method """
OneClassClassifierBase.train(self, data, label)
[docs] def _execute(self, data):
result = RMM2Node._execute(self, data)
label = result.label
prediction = result.prediction + 1
if self.outer_boundary and \
prediction < (-1.0 * self.range + 1) * 0.5:
prediction = (-1.0 * self.range + 1) - prediction
if prediction > 0:
label = self.classes[1]
elif 0 >= prediction:
label = self.classes[0]
return PredictionVector(prediction=prediction,
label=label,
predictor=self)
[docs] def _inc_train(self, data, label):
""" Special wrapper needed to avoid wrong or unknown label
Mostly code copy from train method.
"""
#one vs. REST case
if "REST" in self.classes and not label in self.classes:
label = "REST"
# one vs. one case
if not self.multinomial and len(self.classes) == 2 \
and not label in self.classes:
return
if len(self.classes) == 0:
self.classes.append(label)
self._log("No positive class label given in: %s. Taking now: %s."\
%(self.__class__.__name__, label),
level=logging.ERROR)
if not label == self.classes[0]:
if len(self.classes) == 1:
self.classes.append(label)
self._log("No negative class label given in: %s. Taking now: %s."\
%(self.__class__.__name__, label),
level=logging.WARNING)
return
super(OcRmmNode, self)._inc_train(data, label)
@NoOptimizationParameter("offset_factor")
@BooleanParameter("outer_boundary")
[docs]class OcRmmPerceptronNode(RmmPerceptronNode):
""" Take zero as opposite class for online learning update formula
**References**
========= ==================================================
main
========= ==================================================
author Krell, M. M. and Woehrle, H.
title `New one-class classifiers based on the origin separation approach <http://dx.doi.org/10.1016/j.patrec.2014.11.008>`_
journal Pattern Recognition Letters
publisher Elsevier
doi 10.1016/j.patrec.2014.11.008
year 2015
========= ==================================================
**Exemplary Call**
.. code-block:: yaml
- node : OcRmmPerceptronNode
parameters :
complexity : 1.0
class_labels : ['Target']
range : 4
:input: FeatureVector
:output: PredictionVector
:Author: Mario Michael Krell
:Created: 2014/01/02
"""
[docs] def __init__(self, outer_boundary=False, **kwargs):
if "offset_factor" in kwargs:
kwargs.pop("offset_factor")
RmmPerceptronNode.__init__(self, offset_factor=0, **kwargs)
self.set_permanent_attributes(
b=1, weight=[1, 1], one_class=True, outer_boundary=outer_boundary,
samples="unused")
[docs] def train(self,data,label):
""" Code copy from OneClassClassifierBase """
# print label, type(label), type(self.classes[0]), self.classes, label in self.classes
#one vs. REST case
if "REST" in self.classes and not label in self.classes:
label = "REST"
# one vs. one case
if not self.multinomial and len(self.classes) == 2 \
and not label in self.classes:
return
if len(self.classes)==0:
self.classes.append(label)
self._log("No positive class label given in: %s. Taking now: %s."\
%(self.__class__.__name__,label),
level=logging.ERROR)
if not label==self.classes[0]:
if len(self.classes)==1:
self.classes.append(label)
self._log("No negative class label given in: %s. Taking now: %s."\
%(self.__class__.__name__,label),
level=logging.WARNING)
return
super(OcRmmPerceptronNode, self).train(data, label)
[docs] def _inc_train(self, data, label):
""" Special wrapper needed to avoid wrong or unknown label
Mostly code copy from train method.
"""
#one vs. REST case
if "REST" in self.classes and not label in self.classes:
label = "REST"
# one vs. one case
if not self.multinomial and len(self.classes) == 2 \
and not label in self.classes:
return
if len(self.classes) == 0:
self.classes.append(label)
self._log("No positive class label given in: %s. Taking now: %s."\
%(self.__class__.__name__,label),
level=logging.ERROR)
if not label == self.classes[0]:
if len(self.classes) == 1:
self.classes.append(label)
self._log("No negative class label given in: %s. Taking now: %s."\
%(self.__class__.__name__,label),
level=logging.WARNING)
return
self._train(data, label)
[docs] def _execute(self, data):
result = RmmPerceptronNode._execute(self, data)
label = result.label
prediction = result.prediction + 1
if prediction > 0:
label = self.classes[1]
elif 0 >= prediction > -1.0 * self.range + 1:
label = self.classes[0]
elif -1.0 * self.range + 1 >= prediction and self.outer_boundary:
label = self.classes[1]
prediction += self.range - 1
elif -1.0 * self.range + 1 >= prediction:
label = self.classes[0]
return PredictionVector(prediction=prediction,
label=label,
predictor=self)
@NoOptimizationParameter("squared_loss")
[docs]class L2OcRmmPerceptronNode(OcRmmPerceptronNode):
""" Squared loss variant of the one-class RMM Perceptron
**References**
========= ==================================================
main
========= ==================================================
author Krell, M. M. and Woehrle, H.
title `New one-class classifiers based on the origin separation approach <http://dx.doi.org/10.1016/j.patrec.2014.11.008>`_
journal Pattern Recognition Letters
publisher Elsevier
doi 10.1016/j.patrec.2014.11.008
year 2015
========= ==================================================
.. seealso::
:class:`OcRmmPerceptronNode`
**Exemplary Call**
.. code-block:: yaml
- node : L2OcRmmPerceptronNode
parameters :
complexity : 1.0
class_labels : ['Target']
range : 4
:input: FeatureVector
:output: PredictionVector
:Author: Mario Michael Krell
:Created: 2014/04/28
"""
[docs] def __init__(self, **kwargs):
if "squared_loss" in kwargs:
kwargs.pop("squared_loss")
OcRmmPerceptronNode.__init__(self, squared_loss=True, **kwargs)
@LogUniformParameter("radius", min_value=0, max_value=10000)
@BooleanParameter("radius_opt")
@ChoiceParameter("version", choices=[0, 1, 2])
[docs]class SvddPassiveAggressiveNode(OcRmmPerceptronNode):
""" Support Vector Data Description like Perceptron's suggested by Crammer
**References**
========= ==================================================
main
========= ==================================================
author Crammer, K. and Dekel, O. and Keshet, J. and Shalev-Shwartz, S. and Singer, Y.
title `Online Passive-Aggressive Algorithms <http://dx.doi.org/10.1016/j.patrec.2013.09.018>`_
journal Journal of Machine Learning Research
doi 10.1016/j.patrec.2013.09.018
volume 7
pages 551 - 585
year 2006
========= ==================================================
**Parameters**
:radius: Maximum range parameter allowed for sphere
(*optional, default: 0*)
:radius_opt: Optimize the range parameter as described in *reference*.
If no optimization is used, the radius parameter defines
the used range.
(*optional, default: False*)
:version: Defines the handling of loss:
* 0: hard margin with zero loss on new sample (PA0),
* 1: soft margin with linear loss punishment (PA1),
* 2: soft margin with squared loss punishment (PA2).
For more details refer to the given *reference*.
(*optional, default: 1*)
**Exemplary Call**
.. code-block:: yaml
- node : SvddPassiveAggressive
parameters :
complexity : 1.0
class_labels : ['Target', 'REST']
radius : 2
radius_opt : True
version : 1
:input: FeatureVector
:output: PredictionVector
:Author: Mario Michael Krell
:Created: 2014/04/17
"""
[docs] def __init__(self, radius=0, version=1, radius_opt=False, **kwargs):
OcRmmPerceptronNode.__init__(self, **kwargs)
self.set_permanent_attributes(
radius=float(radius) if not radius_opt else 0.0,
center=None,
version=version,
max_radius=float(radius),)
[docs] def _train(self, data, class_label):
if self.feature_names is None:
try:
self.feature_names = data.feature_names
except AttributeError as e:
warnings.warn("Use a feature generator node before a " +
"classification node.")
raise e
if self.dim is None:
self.dim = data.shape[1]
if class_label not in self.classes and not "REST" in self.classes:
warnings.warn("Please give the expected classes to the classifier! "
+ "%s unknown. "%class_label
+ "Therefore define the variable 'class_labels' in "
+ "your spec file, where you use your classifier. "
+ "For further info look at the node documentation.")
if not(len(self.classes) == 2):
self.classes.append(class_label)
self.set_permanent_attributes(classes=self.classes)
data_array = data.view(numpy.ndarray)
# individual initialization of classification vector
if self.center is None:
self.center = data_array
vector_distance = numpy.linalg.norm(self.center - data_array)
if self.max_radius <= self.radius:
distance = vector_distance
radius_distance = 0
else:
# extend new vector with zero and center with radius distance
# for distance calculation
radius_distance = numpy.sqrt(self.max_radius**2 - self.radius**2)
# norm(a,b)=norm(norm(a),norm(b))
distance = numpy.linalg.norm([vector_distance, radius_distance])
if distance > self.max_radius:
loss = distance - self.max_radius
else:
loss = 0
# no change needed
if loss == 0:
# update_factor = 0
return
# PA0
elif self.version == 0:
update_factor = loss
# PA1
elif self.version == 1:
update_factor = numpy.min([self.complexity, loss])
# PA2
elif self.version == 2:
update_factor = loss / (1.0 + 1.0 / (2.0 * self.complexity))
self.center += update_factor / distance * (data_array - self.center)
# extend new vector with zero and center with radius distance
# and apply same update step
radius_distance += update_factor / distance * (0 - radius_distance)
if radius_distance <= 0:
self.radius = self.max_radius
else: # recalculation of radius from distance
self.radius = numpy.sqrt(self.max_radius**2 - radius_distance**2)
[docs] def _execute(self, data):
data_array = data.view(numpy.ndarray)
if self.center is None:
self.center = data_array
distance = float(numpy.linalg.norm(self.center - data_array))
prediction = distance - self.radius
if prediction > 0 :
label = self.classes[1]
else:
label = self.classes[0]
return PredictionVector(prediction=prediction,
label=label,
predictor=self)
@NoOptimizationParameter("version")
[docs]class UnaryPA0Node(SvddPassiveAggressiveNode):
""" PA0 Node for unary classification
.. seealso::
:class:`SvddPassiveAggressiveNode`
**Exemplary Call**
.. code-block:: yaml
- node : UnaryPA0
parameters :
complexity : 1.0
class_labels : ['Target', 'REST']
radius : 2
radius_opt : True
:input: FeatureVector
:output: PredictionVector
:Author: Mario Michael Krell
:Created: 2014/04/17
"""
[docs] def __init__(self, **kwargs):
kwargs.pop("version", "")
SvddPassiveAggressiveNode.__init__(self, version=0, **kwargs)
@NoOptimizationParameter("version")
[docs]class UnaryPA1Node(SvddPassiveAggressiveNode):
""" PA1 Node for unary classification
.. seealso::
:class:`SvddPassiveAggressiveNode`
**Exemplary Call**
.. code-block:: yaml
- node : UnaryPA1
parameters :
complexity : 1.0
class_labels : ['Target', 'REST']
radius : 2
radius_opt : True
:input: FeatureVector
:output: PredictionVector
:Author: Mario Michael Krell
:Created: 2014/04/17
"""
[docs] def __init__(self, **kwargs):
kwargs.pop("version", "")
SvddPassiveAggressiveNode.__init__(self, version=1, **kwargs)
@NoOptimizationParameter("version")
[docs]class UnaryPA2Node(SvddPassiveAggressiveNode):
""" PA2 Node for unary classification
.. seealso::
:class:`SvddPassiveAggressiveNode`
**Exemplary Call**
.. code-block:: yaml
- node : UnaryPA2
parameters :
complexity : 1.0
class_labels : ['Target', 'REST']
radius : 2
radius_opt : True
:input: FeatureVector
:output: PredictionVector
:Author: Mario Michael Krell
:Created: 2014/04/17
"""
[docs] def __init__(self, **kwargs):
kwargs.pop("version", "")
SvddPassiveAggressiveNode.__init__(self, version=2, **kwargs)
