def trainNormalization(self):
assert len(self.labels) >= 2
if self.type == TYPE_TWOCLASS or self.type == TYPE_MULTICLASS:
# Learn the classes
n_classes = 0
self.class_map = {}
for label in self.labels:
if not self.class_map.has_key(label):
self.class_map[label] = n_classes
n_classes += 1
if self.type == TYPE_MULTICLASS:
assert n_classes >= 2
if self.type == TYPE_TWOCLASS:
assert n_classes == 2
self.class_inv = {}
for key, value in self.class_map.iteritems():
self.class_inv[value] = key
new_labels = []
for each in self.labels:
new_labels.append(self.class_map[each])
self.labels = new_labels
if self.type == TYPE_REGRESSION:
self.reg_mean = mean(self.labels)
self.reg_std = std(self.labels)
new_labels = []
for each in self.labels:
new_labels.append((each - self.reg_mean) / self.reg_std)
self.labels = new_labels
#test length
shape = self.vectors[0].shape
assert len(shape) == 1
for each in self.vectors:
assert shape == each.shape
#crate a data matrix
data = array(self.vectors, 'd')
if self.norm == NORM_AUTO:
self.norm = NORM_VALUE
if data.shape[1] > 128:
self.norm = NORM_PCA
#Setup value normalization
if self.norm == NORM_VALUE:
self.dmean = data.mean(axis=0)
self.dstd = data.std(axis=0)
self.vectors = (data - self.dmean) / self.dstd
elif self.norm == NORM_PCA:
self.pca = PCA()
for vec in self.vectors:
self.pca.addFeature(vec)
if self.pca_basis > 1:
self.pca.train(drop_front=self.pca_drop, number=self.pca_basis)
else:
self.pca.train(drop_front=self.pca_drop, energy=self.pca_basis)
new_vectors = []
for each in self.vectors:
new_vectors.append(self.pca.project(each, whiten=True))
self.vectors = array(new_vectors, 'd')
def trainNormalization(self):
assert len(self.labels) >= 2
if self.type == TYPE_TWOCLASS or self.type == TYPE_MULTICLASS:
# Learn the classes
n_classes = 0
self.class_map = {}
for label in self.labels:
if not self.class_map.has_key(label):
self.class_map[label] = n_classes
n_classes+=1
if self.type == TYPE_MULTICLASS:
assert n_classes >= 2
if self.type == TYPE_TWOCLASS:
assert n_classes == 2
self.class_inv = {}
for key,value in self.class_map.iteritems():
self.class_inv[value] = key
new_labels=[]
for each in self.labels:
new_labels.append(self.class_map[each])
self.labels = new_labels
if self.type == TYPE_REGRESSION:
self.reg_mean = mean(self.labels)
self.reg_std = std(self.labels)
new_labels=[]
for each in self.labels:
new_labels.append((each - self.reg_mean)/self.reg_std)
self.labels = new_labels
#test length
shape = self.vectors[0].shape
assert len(shape) == 1
for each in self.vectors:
assert shape == each.shape
#crate a data matrix
data = array(self.vectors,'d')
if self.norm == NORM_AUTO:
self.norm = NORM_VALUE
if data.shape[1] > 128:
self.norm = NORM_PCA
#Setup value normalization
if self.norm == NORM_VALUE:
self.dmean = data.mean(axis=0)
self.dstd = data.std(axis=0)
self.vectors = (data-self.dmean)/self.dstd
elif self.norm == NORM_PCA:
self.pca = PCA()
for vec in self.vectors:
self.pca.addFeature(vec)
if self.pca_basis > 1:
self.pca.train(drop_front=self.pca_drop,number=self.pca_basis)
else:
self.pca.train(drop_front=self.pca_drop,energy=self.pca_basis)
new_vectors = []
for each in self.vectors:
new_vectors.append(self.pca.project(each,whiten=True))
self.vectors=array(new_vectors,'d')
class VectorClassifier:
##
# Configure some defaults for the classifier value normalizion.
#
# <p>This configures some defalts for the classifier such as the
# type of classifier, and how values are normalized.
def __init__(self,
classifer_type,
normalization=NORM_AUTO,
reg_norm=REG_NORM_VALUE,
pca_basis=0.95,
pca_drop=0):
# Setup basic configuration
self.type = classifer_type
self.norm = normalization
self.reg_norm = reg_norm
self.pca_basis = pca_basis
self.pca_drop = pca_drop
self.labels = []
self.vectors = []
self.vector_length = None
self.reg_mean = 0.0
self.reg_std = 1.0
##
# Learn the range of values that are expected for labels and data.
# Then setup for normalization.
def trainNormalization(self):
assert len(self.labels) >= 2
if self.type == TYPE_TWOCLASS or self.type == TYPE_MULTICLASS:
# Learn the classes
n_classes = 0
self.class_map = {}
for label in self.labels:
if not self.class_map.has_key(label):
self.class_map[label] = n_classes
n_classes += 1
if self.type == TYPE_MULTICLASS:
assert n_classes >= 2
if self.type == TYPE_TWOCLASS:
assert n_classes == 2
self.class_inv = {}
for key, value in self.class_map.iteritems():
self.class_inv[value] = key
new_labels = []
for each in self.labels:
new_labels.append(self.class_map[each])
self.labels = new_labels
if self.type == TYPE_REGRESSION:
self.reg_mean = mean(self.labels)
self.reg_std = std(self.labels)
new_labels = []
for each in self.labels:
new_labels.append((each - self.reg_mean) / self.reg_std)
self.labels = new_labels
#test length
shape = self.vectors[0].shape
assert len(shape) == 1
for each in self.vectors:
assert shape == each.shape
#crate a data matrix
data = array(self.vectors, 'd')
if self.norm == NORM_AUTO:
self.norm = NORM_VALUE
if data.shape[1] > 128:
self.norm = NORM_PCA
#Setup value normalization
if self.norm == NORM_VALUE:
self.dmean = data.mean(axis=0)
self.dstd = data.std(axis=0)
self.vectors = (data - self.dmean) / self.dstd
elif self.norm == NORM_PCA:
self.pca = PCA()
for vec in self.vectors:
self.pca.addFeature(vec)
if self.pca_basis > 1:
self.pca.train(drop_front=self.pca_drop, number=self.pca_basis)
else:
self.pca.train(drop_front=self.pca_drop, energy=self.pca_basis)
new_vectors = []
for each in self.vectors:
new_vectors.append(self.pca.project(each, whiten=True))
self.vectors = array(new_vectors, 'd')
##
# Normalize the values in a data vector to be mean zero.
def normalizeVector(self, data):
if self.norm == NORM_NONE:
return data
elif self.norm == NORM_VALUE:
return (data - self.dmean) / self.dstd
elif self.norm == NORM_PCA:
return self.pca.project(data, whiten=True)
else:
raise NotImplementedError(
"Could not determine nomalization type: " + self.norm)
##
# Add a training sample. Data must be a vector of numbers.
def addTraining(self, label, data, ilog=None):
if self.type == TYPE_REGRESSION:
self.labels.append(float(label))
else:
self.labels.append(label)
if isinstance(data, pv.Image):
data = data.asMatrix2D().flatten()
data = array(data, 'd').flatten()
self.vectors.append(data)
##
# Predict the class or the value for the input data.
#
# <p>This function will perform value normalization and then
# delegate to the subclass to perform classifiaction or
# regression.
def predict(self, data, ilog=None):
if isinstance(data, pv.Image):
data = data.asMatrix2D().flatten()
data = array(data, 'd').flatten()
data = self.normalizeVector(data)
value = self.predictValue(data, ilog=ilog)
if self.type == TYPE_TWOCLASS or self.type == TYPE_MULTICLASS:
return self.invertClass(value)
if self.type == TYPE_REGRESSION:
return self.invertReg(value)
##
# Override this method in subclasses.
# Input should be a numpy array of doubles
#
# If classifer output is int
# If regression output is float
def predictValue(self, data):
raise NotImplementedError("This is an abstract method")
##
# Train the classifer on the training data.
#
# This normalizes the data and the labels, and then passes the
# results to the subclass for training.
def train(self, ilog=None, **kwargs):
self.trainNormalization()
self.trainClassifer(self.labels, self.vectors, ilog=ilog, **kwargs)
# remove training data
del self.labels
del self.vectors
##
# This abstract method should be overridden by subclasses.
#
# <p> This method is called from {@link train}. The vectors and values
# passed to this method will have been normalized. This method is should
# train a classifier or regression algorithm for that normalized data.
#
# <p> Any keyword arguments passed to train will also be passed on to train
# classifier. This could allow variations in training or for verbose
# output.
def trainClassifer(self, labels, vectors, ilog=None, **kwargs):
raise NotImplementedError("This is an abstract method")
##
# Convert a normalized regression value back to the original scale
def invertReg(self, value):
return value * self.reg_std + self.reg_mean
##
# Convert an integer class value back to the original label values.
def invertClass(self, value):
'''Map an integer back into a class label'''
return self.class_inv[value]
class VectorClassifier:
##
# Configure some defaults for the classifier value normalizion.
#
# <p>This configures some defalts for the classifier such as the
# type of classifier, and how values are normalized.
def __init__(self, classifer_type, normalization=NORM_AUTO, reg_norm=REG_NORM_VALUE, pca_basis=0.95, pca_drop=0):
# Setup basic configuration
self.type = classifer_type
self.norm = normalization
self.reg_norm = reg_norm
self.pca_basis = pca_basis
self.pca_drop = pca_drop
self.labels = []
self.vectors = []
self.vector_length = None
self.reg_mean = 0.0
self.reg_std = 1.0
##
# Learn the range of values that are expected for labels and data.
# Then setup for normalization.
def trainNormalization(self):
assert len(self.labels) >= 2
if self.type == TYPE_TWOCLASS or self.type == TYPE_MULTICLASS:
# Learn the classes
n_classes = 0
self.class_map = {}
for label in self.labels:
if not self.class_map.has_key(label):
self.class_map[label] = n_classes
n_classes+=1
if self.type == TYPE_MULTICLASS:
assert n_classes >= 2
if self.type == TYPE_TWOCLASS:
assert n_classes == 2
self.class_inv = {}
for key,value in self.class_map.iteritems():
self.class_inv[value] = key
new_labels=[]
for each in self.labels:
new_labels.append(self.class_map[each])
self.labels = new_labels
if self.type == TYPE_REGRESSION:
self.reg_mean = mean(self.labels)
self.reg_std = std(self.labels)
new_labels=[]
for each in self.labels:
new_labels.append((each - self.reg_mean)/self.reg_std)
self.labels = new_labels
#test length
shape = self.vectors[0].shape
assert len(shape) == 1
for each in self.vectors:
assert shape == each.shape
#crate a data matrix
data = array(self.vectors,'d')
if self.norm == NORM_AUTO:
self.norm = NORM_VALUE
if data.shape[1] > 128:
self.norm = NORM_PCA
#Setup value normalization
if self.norm == NORM_VALUE:
self.dmean = data.mean(axis=0)
self.dstd = data.std(axis=0)
self.vectors = (data-self.dmean)/self.dstd
elif self.norm == NORM_PCA:
self.pca = PCA()
for vec in self.vectors:
self.pca.addFeature(vec)
if self.pca_basis > 1:
self.pca.train(drop_front=self.pca_drop,number=self.pca_basis)
else:
self.pca.train(drop_front=self.pca_drop,energy=self.pca_basis)
new_vectors = []
for each in self.vectors:
new_vectors.append(self.pca.project(each,whiten=True))
self.vectors=array(new_vectors,'d')
##
# Normalize the values in a data vector to be mean zero.
def normalizeVector(self,data):
if self.norm == NORM_NONE:
return data
elif self.norm == NORM_VALUE:
return (data-self.dmean)/self.dstd
elif self.norm == NORM_PCA:
return self.pca.project(data,whiten=True)
else:
raise NotImplementedError("Could not determine nomalization type: "+ self.norm)
##
# Add a training sample. Data must be a vector of numbers.
def addTraining(self,label,data,ilog=None):
if self.type == TYPE_REGRESSION:
self.labels.append(float(label))
else:
self.labels.append(label)
if isinstance(data,pv.Image):
data = data.asMatrix2D().flatten()
data = array(data,'d').flatten()
self.vectors.append(data)
##
# Predict the class or the value for the input data.
#
# <p>This function will perform value normalization and then
# delegate to the subclass to perform classifiaction or
# regression.
def predict(self,data,ilog=None):
if isinstance(data,pv.Image):
data = data.asMatrix2D().flatten()
data = array(data,'d').flatten()
data = self.normalizeVector(data)
value = self.predictValue(data,ilog=ilog)
if self.type == TYPE_TWOCLASS or self.type == TYPE_MULTICLASS:
return self.invertClass(value)
if self.type == TYPE_REGRESSION:
return self.invertReg(value)
##
# Override this method in subclasses.
# Input should be a numpy array of doubles
#
# If classifer output is int
# If regression output is float
def predictValue(self,data):
raise NotImplementedError("This is an abstract method")
##
# Train the classifer on the training data.
#
# This normalizes the data and the labels, and then passes the
# results to the subclass for training.
def train(self,ilog=None,**kwargs):
self.trainNormalization()
self.trainClassifer(self.labels,self.vectors,ilog=ilog,**kwargs)
# remove training data
del self.labels
del self.vectors
##
# This abstract method should be overridden by subclasses.
#
# <p> This method is called from {@link train}. The vectors and values
# passed to this method will have been normalized. This method is should
# train a classifier or regression algorithm for that normalized data.
#
# <p> Any keyword arguments passed to train will also be passed on to train
# classifier. This could allow variations in training or for verbose
# output.
def trainClassifer(self,labels,vectors,ilog=None, **kwargs):
raise NotImplementedError("This is an abstract method")
##
# Convert a normalized regression value back to the original scale
def invertReg(self,value):
return value*self.reg_std + self.reg_mean
##
# Convert an integer class value back to the original label values.
def invertClass(self,value):
'''Map an integer back into a class label'''
return self.class_inv[value]