def selectFeatures(self, _data, *options, **args) :
self.eliminated = []
self.measures = []
cvArgs = {}
import re
rocExp = re.compile(r"roc(?P<rocN>[0-9]+)area")
match = rocExp.match(self.measure)
if match is not None :
measureStr = 'rocNarea'
cvArgs['rocN'] = match.groupdict()['rocN']
else :
measureStr = self.measure
data = _data.__class__(_data, deepcopy = True)
for i in range(self.targetNumFeatures, _data.numFeatures) :
maxScore = 0
# loop over the CURRENT features
for feature in range(data.numFeatures) :
featureName = data.featureID[feature]
data.eliminateFeatures([feature])
res = self.classifier.stratifiedCV(data, **cvArgs)
score = getattr(res, measureStr)
if score > maxScore :
maxScore = score
bestFeatureName = featureName
data = _data.__class__(_data, deepcopy = True)
data.eliminateFeatures(data.featureNames2IDs(self.eliminated))
data = _data.__class__(_data, deepcopy = True)
self.eliminated.append(bestFeatureName)
data.eliminateFeatures(data.featureNames2IDs(self.eliminated))
self.measures.append(maxScore)
return misc.setminus(range(_data.numFeatures),
_data.featureNames2IDs(self.eliminated))
def selectFeatures(self, _data, *options, **args):
self.eliminated = []
self.measures = []
cvArgs = {}
import re
rocExp = re.compile(r"roc(?P<rocN>[0-9]+)area")
match = rocExp.match(self.measure)
if match is not None:
measureStr = 'rocNarea'
cvArgs['rocN'] = match.groupdict()['rocN']
else:
measureStr = self.measure
data = _data.__class__(_data, deepcopy=True)
for i in range(self.targetNumFeatures, _data.numFeatures):
maxScore = 0
# loop over the CURRENT features
for feature in range(data.numFeatures):
featureName = data.featureID[feature]
data.eliminateFeatures([feature])
res = self.classifier.stratifiedCV(data, **cvArgs)
score = getattr(res, measureStr)
if score > maxScore:
maxScore = score
bestFeatureName = featureName
data = _data.__class__(_data, deepcopy=True)
data.eliminateFeatures(data.featureNames2IDs(self.eliminated))
data = _data.__class__(_data, deepcopy=True)
self.eliminated.append(bestFeatureName)
data.eliminateFeatures(data.featureNames2IDs(self.eliminated))
self.measures.append(maxScore)
return misc.setminus(range(_data.numFeatures),
_data.featureNames2IDs(self.eliminated))
def makeFolds(data, numFolds, datasetName, directory = '.') :
'''split a dataset into several folds and save the training and testing
data of each fold as a separate dataset
data - a dataset instance
numfolds - number of folds into which to split the data
datasetName - string to use for the file names
directory - the directory into which to deposit the files
'''
perm = range(len(data))
random.shuffle(perm)
foldSize = len(data) / numFolds
for fold in range(numFolds) :
if fold < numFolds-1:
testingPatterns = perm[foldSize * fold : foldSize * (fold + 1)]
else:
testingPatterns = perm[foldSize * fold : len(data)]
trainingPatterns = misc.setminus(range(len(data)), testingPatterns)
trainingData = data.__class__(data, patterns = trainingPatterns)
testingData = data.__class__(data, patterns = testingPatterns)
testingDataName = os.path.join(directory, datasetName + 'Testing' + str(fold) + '.data')
testingData.save(testingDataName)
trainingDataName = os.path.join(directory, datasetName + 'Training' + str(fold) + '.data')
trainingData.save(trainingDataName)
def makeFolds(data, numFolds, datasetName, directory='.'):
'''split a dataset into several folds and save the training and testing
data of each fold as a separate dataset
data - a dataset instance
numfolds - number of folds into which to split the data
datasetName - string to use for the file names
directory - the directory into which to deposit the files
'''
perm = range(len(data))
random.shuffle(perm)
foldSize = len(data) / numFolds
for fold in range(numFolds):
if fold < numFolds - 1:
testingPatterns = perm[foldSize * fold:foldSize * (fold + 1)]
else:
testingPatterns = perm[foldSize * fold:len(data)]
trainingPatterns = misc.setminus(range(len(data)), testingPatterns)
trainingData = data.__class__(data, patterns=trainingPatterns)
testingData = data.__class__(data, patterns=testingPatterns)
testingDataName = os.path.join(
directory, datasetName + 'Testing' + str(fold) + '.data')
testingData.save(testingDataName)
trainingDataName = os.path.join(
directory, datasetName + 'Training' + str(fold) + '.data')
trainingData.save(trainingDataName)
def keepFeatures(self, features) :
"""eliminate all but the give list of features
INPUT:
features - a list of features to eliminate; these are either numbers
between 0 and numFeatures-1 (indices of features, not their IDs) or
featureIDs
"""
if type(features[0]) == type('') :
features = self.featureNames2IDs(features)
self.eliminateFeatures(misc.setminus(range(self.numFeatures), features))
def keepFeatures(self, features):
"""eliminate all but the give list of features
INPUT:
features - a list of features to eliminate; these are either numbers
between 0 and numFeatures-1 (indices of features, not their IDs) or
featureIDs
"""
if type(features[0]) == type(''):
features = self.featureNames2IDs(features)
self.eliminateFeatures(misc.setminus(range(self.numFeatures),
features))
def selectFeatures(self, data, targetClass=None, otherClass = None, *options, **args) :
s = self.featureScore.score(data, targetClass, otherClass, **args)
if self.mode == "byNum" :
featuresToEliminate = numpy.argsort(s)\
[:data.numFeatures - self.numFeatures]
elif self.mode == "byThreshold" :
featuresToEliminate = numpy.nonzero(numpy.less(s, self.threshold))[0]
elif self.mode == "bySignificance" :
t = self.significanceThreshold(data)
self.thresholds = t
featuresToEliminate = numpy.nonzero(numpy.less(s, t))[0]
else :
raise ValueError, 'unknown elimination mode in filter'
return misc.setminus(range(data.numFeatures), featuresToEliminate)
def eliminateFeatures(self, featureList):
"""eliminate a list of features from a dataset
Input:
featureList - a list of features to eliminate; these are numbers
between 0 and numFeatures-1 (indices of features, not their IDs)"""
#if len(featureList) == 0 : return
#if type(featureList[0]) == type('') :
# featureList = self.featureNames2IDs(features)
featuresToTake = misc.setminus(range(self.numFeatures), featureList)
featuresToTake.sort()
self.featureID = [self.featureID[i] for i in featuresToTake]
#self.featureKey = [self.featureKey[i] for i in featuresToTake]
#self.featureKeyDict = {}
#for i in range(len(self.featureKey)) :
# self.featureKeyDict[self.featureKey[i]] = i
self.X = numpy.take(self.X, featuresToTake, 1)
def eliminateFeatures(self, featureList) :
"""eliminate a list of features from a dataset
Input:
featureList - a list of features to eliminate; these are numbers
between 0 and numFeatures-1 (indices of features, not their IDs)"""
if len(featureList) == 0 : return
if type(featureList[0]) == type('') :
featureList = self.featureNames2IDs(features)
featuresToTake = misc.setminus(range(self.numFeatures), featureList)
featuresToTake.sort()
self.featureID = [self.featureID[i] for i in featuresToTake]
self.featureKey = [self.featureKey[i] for i in featuresToTake]
self.featureKeyDict = {}
for i in range(len(self.featureKey)) :
self.featureKeyDict[self.featureKey[i]] = i
self.X = numpy.take(self.X, featuresToTake, 1)
def cv(classifier, data, numFolds=5, **args):
"""perform k-fold cross validation
:Parameters:
- `classifier` - a classifier template
- `data` - a dataset
- `numFolds` - number of cross validation folds (default = 5)
:Returns:
a Results object.
:Keywords:
- `numFolds` - number of cross validation folds (default = 5)
- `seed` - random number generator seed
- `foldsToPerform` - number of folds to actually perform (in case you're doing
n fold CV, and want to save time, and only do some of the folds)
"""
if 'numFolds' in args:
numFolds = args['numFolds']
if 'seed' in args:
random.seed(args['seed'])
foldsToPerform = numFolds
if 'foldsToPerform' in args:
foldsToPerform = args['foldsToPerform']
if foldsToPerform > numFolds:
raise ValueError, 'foldsToPerform > numFolds'
perm = range(len(data))
random.shuffle(perm)
foldSize = len(data) / numFolds
trainingPatterns = []
testingPatterns = []
for fold in range(foldsToPerform):
if fold < numFolds - 1:
testingPatterns.append(perm[foldSize * fold:foldSize * (fold + 1)])
else:
testingPatterns.append(perm[foldSize * fold:len(data)])
trainingPatterns.append(
misc.setminus(range(len(data)), testingPatterns[-1]))
return cvFromFolds(classifier, data, trainingPatterns, testingPatterns,
**args)
def cv(classifier, data, numFolds = 5, **args) :
"""perform k-fold cross validation
:Parameters:
- `classifier` - a classifier template
- `data` - a dataset
- `numFolds` - number of cross validation folds (default = 5)
:Returns:
a Results object.
:Keywords:
- `numFolds` - number of cross validation folds (default = 5)
- `seed` - random number generator seed
- `foldsToPerform` - number of folds to actually perform (in case you're doing
n fold CV, and want to save time, and only do some of the folds)
"""
if 'numFolds' in args :
numFolds = args['numFolds']
if 'seed' in args :
random.seed(args['seed'])
foldsToPerform = numFolds
if 'foldsToPerform' in args :
foldsToPerform = args['foldsToPerform']
if foldsToPerform > numFolds :
raise ValueError, 'foldsToPerform > numFolds'
perm = range(len(data))
random.shuffle(perm)
foldSize = len(data) / numFolds
trainingPatterns = []
testingPatterns = []
for fold in range(foldsToPerform) :
if fold < numFolds-1:
testingPatterns.append(perm[foldSize * fold : foldSize * (fold + 1)])
else:
testingPatterns.append(perm[foldSize * fold : len(data)])
trainingPatterns.append(misc.setminus(range(len(data)),
testingPatterns[-1]))
return cvFromFolds(classifier, data, trainingPatterns, testingPatterns, **args)
def loo(classifier, data, **args) :
"""perform Leave One Out
:Returns:
a results object
USAGE: loo(classifier, data)
"""
looResults = classifier.resultsObject()
args['stats'] = False
for i in range(len(data)) :
trainingPatterns = misc.setminus(range(len(data)), [i])
looResults.extend(
classifier.trainTest(data, trainingPatterns, [i], **args))
looResults.computeStats()
return looResults
def loo(classifier, data, **args):
"""perform Leave One Out
:Returns:
a results object
USAGE: loo(classifier, data)
"""
looResults = classifier.resultsObject()
args['stats'] = False
for i in range(len(data)):
trainingPatterns = misc.setminus(range(len(data)), [i])
looResults.extend(
classifier.trainTest(data, trainingPatterns, [i], **args))
looResults.computeStats()
return looResults
def selectFeatures(self,
data,
targetClass=None,
otherClass=None,
*options,
**args):
s = self.featureScore.score(data, targetClass, otherClass, **args)
if self.mode == "byNum":
featuresToEliminate = numpy.argsort(s)\
[:data.numFeatures - self.numFeatures]
elif self.mode == "byThreshold":
featuresToEliminate = numpy.nonzero(numpy.less(s,
self.threshold))[0]
elif self.mode == "bySignificance":
t = self.significanceThreshold(data)
self.thresholds = t
featuresToEliminate = numpy.nonzero(numpy.less(s, t))[0]
else:
raise ValueError, 'unknown elimination mode in filter'
return misc.setminus(range(data.numFeatures), featuresToEliminate)
def split(data, fraction, **args):
"""
split a dataset into training and test sets.
randomly splits a dataset into two datasets whose sizes are determined
by the 'fraction' parameter (the first dataset will contain that fraction
of the examples).
for example:
train, test = split(data, 0.7)
will split the data -- 70% for training and 30% for test
:Parameters:
- `data` - a dataset object
- `fraction` - the fraction of the examples to put in the first split
:Keywords:
- `stratified` - whether to perform stratified splitting, i.e. whether to
keep the class ratio in the two datasets [default: True]
- `seed` - random number generator seed
- `indicesOnly` - if this flag is set, the indices of the two splits are
returned instead of the datasets [default: False]
"""
if 'seed' in args:
seed = args['seed']
rand = random.Random(seed)
else:
rand = random.Random()
indicesOnly = False
if 'indicesOnly' in args:
indicesOnly = args['indicesOnly']
if data.__class__.__name__ == 'Labels':
labels = data
else:
labels = data.labels
sampleSize = int(len(data) * fraction)
stratified = True
if 'stratified' in args:
stratified = args['stratified']
if stratified:
patterns = []
for i in range(labels.numClasses):
if i < labels.numClasses - 1:
numToSample = int(fraction * labels.classSize[i])
else:
numToSample = sampleSize - len(patterns)
I = labels.classes[i][:]
rand.shuffle(I)
patterns.extend(I[:numToSample])
else:
I = range(len(data))
rand.shuffle(I)
patterns = I[:sampleSize]
patterns.sort()
if not indicesOnly:
return (data.__class__(data, patterns=patterns),
data.__class__(data,
patterns=misc.setminus(range(len(data)),
patterns)))
else:
return patterns, misc.setminus(range(len(data)), patterns)
def split(data, fraction, **args) :
"""
split a dataset into training and test sets.
randomly splits a dataset into two datasets whose sizes are determined
by the 'fraction' parameter (the first dataset will contain that fraction
of the examples).
for example:
train, test = split(data, 0.7)
will split the data -- 70% for training and 30% for test
:Parameters:
- `data` - a dataset object
- `fraction` - the fraction of the examples to put in the first split
:Keywords:
- `stratified` - whether to perform stratified splitting, i.e. whether to
keep the class ratio in the two datasets [default: True]
- `seed` - random number generator seed
- `indicesOnly` - if this flag is set, the indices of the two splits are
returned instead of the datasets [default: False]
"""
if 'seed' in args :
seed = args['seed']
rand = random.Random(seed)
else :
rand = random.Random()
indicesOnly = False
if 'indicesOnly' in args :
indicesOnly = args['indicesOnly']
if data.__class__.__name__ == 'Labels' :
labels = data
else :
labels = data.labels
sampleSize = int(len(data) * fraction)
stratified = True
if 'stratified' in args :
stratified = args['stratified']
if stratified :
patterns = []
for i in range(labels.numClasses) :
if i < labels.numClasses - 1 :
numToSample = int(fraction * labels.classSize[i])
else :
numToSample = sampleSize - len(patterns)
I = labels.classes[i][:]
rand.shuffle(I)
patterns.extend(I[:numToSample])
else :
I = range(len(data))
rand.shuffle(I)
patterns = I[:sampleSize]
patterns.sort()
if not indicesOnly :
return (data.__class__(data, patterns = patterns),
data.__class__(data, patterns = misc.setminus(range(len(data)), patterns) ) )
else :
return patterns, misc.setminus(range(len(data)), patterns)