def test_split_classifier(self):
ds = self.data_bin_1
clf = SplitClassifier(
clf=SameSignClassifier(),
splitter=NFoldSplitter(1),
enable_ca=['confusion', 'training_confusion', 'feature_ids'])
clf.train(ds) # train the beast
error = clf.ca.confusion.error
tr_error = clf.ca.training_confusion.error
clf2 = clf.clone()
cv = CrossValidatedTransferError(
TransferError(clf2),
NFoldSplitter(),
postproc=mean_sample(),
enable_ca=['confusion', 'training_confusion'])
cverror = cv(ds).samples.squeeze()
tr_cverror = cv.ca.training_confusion.error
self.failUnlessEqual(
error,
cverror,
msg="We should get the same error using split classifier as"
" using CrossValidatedTransferError. Got %s and %s" %
(error, cverror))
self.failUnlessEqual(
tr_error,
tr_cverror,
msg="We should get the same training error using split classifier as"
" using CrossValidatedTransferError. Got %s and %s" %
(tr_error, tr_cverror))
self.failUnlessEqual(clf.ca.confusion.percent_correct,
100,
msg="Dummy clf should train perfectly")
self.failUnlessEqual(len(clf.ca.confusion.sets),
len(ds.UC),
msg="Should have 1 confusion per each split")
self.failUnlessEqual(
len(clf.clfs),
len(ds.UC),
msg="Should have number of classifiers equal # of epochs")
self.failUnlessEqual(clf.predict(ds.samples),
list(ds.targets),
msg="Should classify correctly")
# feature_ids must be list of lists, and since it is not
# feature-selecting classifier used - we expect all features
# to be utilized
# NOT ANYMORE -- for BoostedClassifier we have now union of all
# used features across slave classifiers. That makes
# semantics clear. If you need to get deeper -- use upcoming
# harvesting facility ;-)
# self.failUnlessEqual(len(clf.feature_ids), len(ds.uniquechunks))
# self.failUnless(np.array([len(ids)==ds.nfeatures
# for ids in clf.feature_ids]).all())
# Just check if we get it at all ;-)
summary = clf.summary()
def test_split_classifier(self):
ds = self.data_bin_1
clf = SplitClassifier(clf=SameSignClassifier(),
splitter=NFoldSplitter(1),
enable_ca=['confusion', 'training_confusion',
'feature_ids'])
clf.train(ds) # train the beast
error = clf.ca.confusion.error
tr_error = clf.ca.training_confusion.error
clf2 = clf.clone()
cv = CrossValidatedTransferError(
TransferError(clf2),
NFoldSplitter(),
postproc=mean_sample(),
enable_ca=['confusion', 'training_confusion'])
cverror = cv(ds).samples.squeeze()
tr_cverror = cv.ca.training_confusion.error
self.failUnlessEqual(error, cverror,
msg="We should get the same error using split classifier as"
" using CrossValidatedTransferError. Got %s and %s"
% (error, cverror))
self.failUnlessEqual(tr_error, tr_cverror,
msg="We should get the same training error using split classifier as"
" using CrossValidatedTransferError. Got %s and %s"
% (tr_error, tr_cverror))
self.failUnlessEqual(clf.ca.confusion.percent_correct,
100,
msg="Dummy clf should train perfectly")
self.failUnlessEqual(len(clf.ca.confusion.sets),
len(ds.UC),
msg="Should have 1 confusion per each split")
self.failUnlessEqual(len(clf.clfs), len(ds.UC),
msg="Should have number of classifiers equal # of epochs")
self.failUnlessEqual(clf.predict(ds.samples), list(ds.targets),
msg="Should classify correctly")
# feature_ids must be list of lists, and since it is not
# feature-selecting classifier used - we expect all features
# to be utilized
# NOT ANYMORE -- for BoostedClassifier we have now union of all
# used features across slave classifiers. That makes
# semantics clear. If you need to get deeper -- use upcoming
# harvesting facility ;-)
# self.failUnlessEqual(len(clf.feature_ids), len(ds.uniquechunks))
# self.failUnless(np.array([len(ids)==ds.nfeatures
# for ids in clf.feature_ids]).all())
# Just check if we get it at all ;-)
summary = clf.summary()
def testRegressions(self, regr):
"""Simple tests on regressions
"""
ds = datasets['chirp_linear']
cve = CrossValidatedTransferError(
TransferError(regr, CorrErrorFx()),
splitter=NFoldSplitter(),
enable_states=['training_confusion', 'confusion'])
corr = cve(ds)
self.failUnless(corr == cve.confusion.stats['CCe'])
splitregr = SplitClassifier(regr,
splitter=OddEvenSplitter(),
enable_states=['training_confusion', 'confusion'])
splitregr.train(ds)
split_corr = splitregr.confusion.stats['CCe']
split_corr_tr = splitregr.training_confusion.stats['CCe']
for confusion, error in ((cve.confusion, corr),
(splitregr.confusion, split_corr),
(splitregr.training_confusion, split_corr_tr),
):
#TODO: test confusion statistics
# Part of it for now -- CCe
for conf in confusion.summaries:
stats = conf.stats
self.failUnless(stats['CCe'] < 0.5)
self.failUnlessEqual(stats['CCe'], stats['Summary CCe'])
s0 = confusion.asstring(short=True)
s1 = confusion.asstring(short=False)
for s in [s0, s1]:
self.failUnless(len(s) > 10,
msg="We should get some string representation "
"of regression summary. Got %s" % s)
self.failUnless(error < 0.2,
msg="Regressions should perform well on a simple "
"dataset. Got correlation error of %s " % error)
# Test access to summary statistics
# YOH: lets start making testing more reliable.
# p-value for such accident to have is verrrry tiny,
# so if regression works -- it better has at least 0.5 ;)
# otherwise fix it! ;)
#if cfg.getboolean('tests', 'labile', default='yes'):
self.failUnless(confusion.stats['CCe'] < 0.5)
split_predictions = splitregr.predict(ds.samples) # just to check if it works fine
def test_regressions(self, regr):
"""Simple tests on regressions
"""
ds = datasets['chirp_linear']
# we want numeric labels to maintain the previous behavior, especially
# since we deal with regressions here
ds.sa.targets = AttributeMap().to_numeric(ds.targets)
cve = CrossValidatedTransferError(
TransferError(regr),
splitter=NFoldSplitter(),
postproc=mean_sample(),
enable_ca=['training_confusion', 'confusion'])
# check the default
self.failUnless(isinstance(cve.transerror.errorfx,
CorrErrorFx))
corr = np.asscalar(cve(ds).samples)
# Our CorrErrorFx should never return NaN
self.failUnless(not np.isnan(corr))
self.failUnless(corr == cve.ca.confusion.stats['CCe'])
splitregr = SplitClassifier(
regr, splitter=OddEvenSplitter(),
enable_ca=['training_confusion', 'confusion'])
splitregr.train(ds)
split_corr = splitregr.ca.confusion.stats['CCe']
split_corr_tr = splitregr.ca.training_confusion.stats['CCe']
for confusion, error in (
(cve.ca.confusion, corr),
(splitregr.ca.confusion, split_corr),
(splitregr.ca.training_confusion, split_corr_tr),
):
#TODO: test confusion statistics
# Part of it for now -- CCe
for conf in confusion.summaries:
stats = conf.stats
if cfg.getboolean('tests', 'labile', default='yes'):
self.failUnless(stats['CCe'] < 0.5)
self.failUnlessEqual(stats['CCe'], stats['Summary CCe'])
s0 = confusion.as_string(short=True)
s1 = confusion.as_string(short=False)
for s in [s0, s1]:
self.failUnless(len(s) > 10,
msg="We should get some string representation "
"of regression summary. Got %s" % s)
if cfg.getboolean('tests', 'labile', default='yes'):
self.failUnless(error < 0.2,
msg="Regressions should perform well on a simple "
"dataset. Got correlation error of %s " % error)
# Test access to summary statistics
# YOH: lets start making testing more reliable.
# p-value for such accident to have is verrrry tiny,
# so if regression works -- it better has at least 0.5 ;)
# otherwise fix it! ;)
# YOH: not now -- issues with libsvr in SG and linear kernel
if cfg.getboolean('tests', 'labile', default='yes'):
self.failUnless(confusion.stats['CCe'] < 0.5)
# just to check if it works fine
split_predictions = splitregr.predict(ds.samples)
def test_regressions(self, regr):
"""Simple tests on regressions
"""
ds = datasets['chirp_linear']
# we want numeric labels to maintain the previous behavior, especially
# since we deal with regressions here
ds.sa.targets = AttributeMap().to_numeric(ds.targets)
cve = CrossValidatedTransferError(
TransferError(regr),
splitter=NFoldSplitter(),
postproc=mean_sample(),
enable_ca=['training_confusion', 'confusion'])
# check the default
self.failUnless(isinstance(cve.transerror.errorfx, CorrErrorFx))
corr = np.asscalar(cve(ds).samples)
# Our CorrErrorFx should never return NaN
self.failUnless(not np.isnan(corr))
self.failUnless(corr == cve.ca.confusion.stats['CCe'])
splitregr = SplitClassifier(
regr,
splitter=OddEvenSplitter(),
enable_ca=['training_confusion', 'confusion'])
splitregr.train(ds)
split_corr = splitregr.ca.confusion.stats['CCe']
split_corr_tr = splitregr.ca.training_confusion.stats['CCe']
for confusion, error in (
(cve.ca.confusion, corr),
(splitregr.ca.confusion, split_corr),
(splitregr.ca.training_confusion, split_corr_tr),
):
#TODO: test confusion statistics
# Part of it for now -- CCe
for conf in confusion.summaries:
stats = conf.stats
if cfg.getboolean('tests', 'labile', default='yes'):
self.failUnless(stats['CCe'] < 0.5)
self.failUnlessEqual(stats['CCe'], stats['Summary CCe'])
s0 = confusion.as_string(short=True)
s1 = confusion.as_string(short=False)
for s in [s0, s1]:
self.failUnless(len(s) > 10,
msg="We should get some string representation "
"of regression summary. Got %s" % s)
if cfg.getboolean('tests', 'labile', default='yes'):
self.failUnless(
error < 0.2,
msg="Regressions should perform well on a simple "
"dataset. Got correlation error of %s " % error)
# Test access to summary statistics
# YOH: lets start making testing more reliable.
# p-value for such accident to have is verrrry tiny,
# so if regression works -- it better has at least 0.5 ;)
# otherwise fix it! ;)
# YOH: not now -- issues with libsvr in SG and linear kernel
if cfg.getboolean('tests', 'labile', default='yes'):
self.failUnless(confusion.stats['CCe'] < 0.5)
# just to check if it works fine
split_predictions = splitregr.predict(ds.samples)