def test_sensitivity_based_feature_selection(self, clf):
# sensitivity analyser and transfer error quantifier use the SAME clf!
sens_ana = clf.get_sensitivity_analyzer(postproc=maxofabs_sample())
# of features to remove
Nremove = 2
# because the clf is already trained when computing the sensitivity
# map, prevent retraining for transfer error calculation
# Use absolute of the svm weights as sensitivity
fe = SensitivityBasedFeatureSelection(
sens_ana,
feature_selector=FixedNElementTailSelector(2),
enable_ca=["sensitivity", "selected_ids"])
data = self.get_data()
data_nfeatures = data.nfeatures
fe.train(data)
resds = fe(data)
# fail if orig datasets are changed
self.assertTrue(data.nfeatures == data_nfeatures)
# silly check if nfeatures got a single one removed
self.assertEqual(data.nfeatures,
resds.nfeatures + Nremove,
msg="We had to remove just a single feature")
self.assertEqual(
fe.ca.sensitivity.nfeatures,
data_nfeatures,
msg="Sensitivity have to have # of features equal to original")
def __repr__(self, prefixes=[]):
return super(SplitRFE, self).__repr__(
prefixes=prefixes
+ _repr_attrs(self, ['lrn', 'partitioner'])
+ _repr_attrs(self, ['errorfx'], default=mean_mismatch_error)
+ _repr_attrs(self, ['analyzer_postproc'], default=maxofabs_sample())
)
def test_sensitivity_based_feature_selection(self, clf):
# sensitivity analyser and transfer error quantifier use the SAME clf!
sens_ana = clf.get_sensitivity_analyzer(postproc=maxofabs_sample())
# of features to remove
Nremove = 2
# because the clf is already trained when computing the sensitivity
# map, prevent retraining for transfer error calculation
# Use absolute of the svm weights as sensitivity
fe = SensitivityBasedFeatureSelection(sens_ana,
feature_selector=FixedNElementTailSelector(2),
enable_ca=["sensitivity", "selected_ids"])
data = self.get_data()
data_nfeatures = data.nfeatures
fe.train(data)
resds = fe(data)
# fail if orig datasets are changed
self.assertTrue(data.nfeatures == data_nfeatures)
# silly check if nfeatures got a single one removed
self.assertEqual(data.nfeatures, resds.nfeatures+Nremove,
msg="We had to remove just a single feature")
self.assertEqual(fe.ca.sensitivity.nfeatures, data_nfeatures,
msg="Sensitivity have to have # of features equal to original")
def __init__(
self,
lrn,
partitioner,
fselector,
errorfx=mean_mismatch_error,
analyzer_postproc=maxofabs_sample(),
# callback?
**kwargs):
"""
Parameters
----------
lrn : Learner
Learner with a sensitivity analyzer which will be used both
for the sensitivity analysis and transfer error estimation
partitioner : Partitioner
Used to generate cross-validation partitions for cross-validation
to deduce optimal number of features to maintain
fselector : Functor
Given a sensitivity map it has to return the ids of those
features that should be kept.
errorfx : func, optional
Functor to use for estimation of cross-validation error
analyzer_postproc : func, optional
Function to provide to the sensitivity analyzer as postproc
"""
# Initialize itself preparing for the 2nd invocation
# with determined number of nfeatures_min
fmeasure = lrn.get_sensitivity_analyzer(postproc=analyzer_postproc)
RFE.__init__(self,
fmeasure,
None,
Repeater(2),
fselector=fselector,
bestdetector=None,
train_pmeasure=False,
stopping_criterion=None,
**kwargs)
self._lrn = lrn # should not be modified, thus _
self.partitioner = partitioner
self.errorfx = errorfx
self.analyzer_postproc = analyzer_postproc
def __init__(self, lrn, partitioner,
fselector,
errorfx=mean_mismatch_error,
analyzer_postproc=maxofabs_sample(),
# callback?
**kwargs):
"""
Parameters
----------
lrn : Learner
Learner with a sensitivity analyzer which will be used both
for the sensitivity analysis and transfer error estimation
partitioner : Partitioner
Used to generate cross-validation partitions for cross-validation
to deduce optimal number of features to maintain
fselector : Functor
Given a sensitivity map it has to return the ids of those
features that should be kept.
errorfx : func, optional
Functor to use for estimation of cross-validation error
analyzer_postproc : func, optional
Function to provide to the sensitivity analyzer as postproc
"""
# Initialize itself preparing for the 2nd invocation
# with determined number of nfeatures_min
fmeasure = lrn.get_sensitivity_analyzer(postproc=analyzer_postproc)
RFE.__init__(self,
fmeasure,
None,
Repeater(2),
fselector=fselector,
bestdetector=None,
train_pmeasure=False,
stopping_criterion=None,
**kwargs)
self._lrn = lrn # should not be modified, thus _
self.partitioner = partitioner
self.errorfx = errorfx
self.analyzer_postproc = analyzer_postproc
def __init__(self, lrn, partitioner,
fselector,
errorfx=mean_mismatch_error,
fmeasure_postproc=None,
fmeasure=None,
nproc=1,
# callback?
**kwargs):
"""
Parameters
----------
lrn : Learner
Learner with a sensitivity analyzer which will be used both
for the sensitivity analysis and transfer error estimation
partitioner : Partitioner
Used to generate cross-validation partitions for cross-validation
to deduce optimal number of features to maintain
fselector : Functor
Given a sensitivity map it has to return the ids of those
features that should be kept.
errorfx : func, optional
Functor to use for estimation of cross-validation error
fmeasure_postproc : func, optional
Function to provide to the sensitivity analyzer as postproc. If no
fmeasure is provided and classifier sensitivity is used, then
maxofabs_sample() would be used for this postproc, unless other
value is provided
fmeasure : Function, optional
Featurewise measure. If None was provided, lrn's sensitivity
analyzer will be used.
"""
# Initialize itself preparing for the 2nd invocation
# with determined number of nfeatures_min
# TODO: move this into _train since better not to assign anything here
# to avoid possible problems with copies needing to deal with the same
# lrn... but then we might like again to reconsider delegation instead
# of subclassing here....
if fmeasure is None:
if __debug__:
debug('RFE', 'fmeasure was not provided, will be using the '
'sensitivity analyzer for %s' % lrn)
fmeasure = lrn.get_sensitivity_analyzer(
postproc=fmeasure_postproc if fmeasure_postproc is not None
else maxofabs_sample())
train_pmeasure = False
else:
assert fmeasure_postproc is None, "There should be no explicit " \
"fmeasure_postproc when fmeasure is specified"
# if user provided explicit value -- use it! otherwise, we do want
# to train an arbitrary fmeasure
train_pmeasure = kwargs.pop('train_pmeasure', True)
RFE.__init__(self,
fmeasure,
None,
Repeater(2),
fselector=fselector,
bestdetector=None,
train_pmeasure=train_pmeasure,
stopping_criterion=None,
**kwargs)
self._lrn = lrn # should not be modified, thus _
self.partitioner = partitioner
self.errorfx = errorfx
self.fmeasure_postproc = fmeasure_postproc
self.nproc = nproc
def test_analyzer_with_split_classifier(self, clfds):
"""Test analyzers in split classifier
"""
clf, ds = clfds # unroll the tuple
# We need to skip some LARSes here
_sclf = str(clf)
if 'LARS(' in _sclf and "type='stepwise'" in _sclf:
# ADD KnownToFail thingie from NiPy
return
# To don't waste too much time testing lets limit to 3 splits
nsplits = 3
partitioner = NFoldPartitioner(count=nsplits)
mclf = SplitClassifier(clf=clf,
partitioner=partitioner,
enable_ca=['training_stats', 'stats'])
sana = mclf.get_sensitivity_analyzer( # postproc=absolute_features(),
pass_attr=['fa.nonbogus_targets'],
enable_ca=["sensitivities"])
ulabels = ds.uniquetargets
nlabels = len(ulabels)
# Can't rely on splitcfg since count-limit is done in __call__
assert (nsplits == len(list(partitioner.generate(ds))))
sens = sana(ds)
assert ('nonbogus_targets' in sens.fa) # were they passsed?
# TODO: those few do not expose biases
if not len(set(clf.__tags__).intersection(('lars', 'glmnet', 'gpr'))):
assert ('biases' in sens.sa)
# print sens.sa.biases
# It should return either ...
# nlabels * nsplits
req_nsamples = [nlabels * nsplits]
if nlabels == 2:
# A single sensitivity in case of binary
req_nsamples += [nsplits]
else:
# and for pairs in case of multiclass
req_nsamples += [(nlabels * (nlabels - 1) / 2) * nsplits]
# and for 1-vs-1 embedded within Multiclass operating on
# pairs (e.g. SMLR)
req_nsamples += [req_nsamples[-1] * 2]
# Also for regression_based -- they can do multiclass
# but only 1 sensitivity is provided
if 'regression_based' in clf.__tags__:
req_nsamples += [nsplits]
# # of features should correspond
self.assertEqual(sens.shape[1], ds.nfeatures)
# # of samples/sensitivities should also be reasonable
self.assertTrue(sens.shape[0] in req_nsamples)
# Check if labels are present
self.assertTrue('splits' in sens.sa)
self.assertTrue('targets' in sens.sa)
# should be 1D -- otherwise dtype object
self.assertTrue(sens.sa.targets.ndim == 1)
sens_ulabels = sens.sa['targets'].unique
# Some labels might be pairs(tuples) so ndarray would be of
# dtype object and we would need to get them all
if sens_ulabels.dtype is np.dtype('object'):
sens_ulabels = np.unique(
reduce(lambda x, y: x + y, [list(x) for x in sens_ulabels]))
assert_array_equal(sens_ulabels, ds.sa['targets'].unique)
errors = [x.percent_correct for x in sana.clf.ca.stats.matrices]
# lets go through all sensitivities and see if we selected the right
# features
#if 'meta' in clf.__tags__ and len(sens.samples[0].nonzero()[0])<2:
if '5%' in clf.descr \
or (nlabels > 2 and 'regression_based' in clf.__tags__):
# Some meta classifiers (5% of ANOVA) are too harsh ;-)
# if we get less than 2 features with on-zero sensitivities we
# cannot really test
# Also -- regression based classifiers performance for multiclass
# is expected to suck in general
return
if cfg.getboolean('tests', 'labile', default='yes'):
for conf_matrix in [sana.clf.ca.training_stats] \
+ sana.clf.ca.stats.matrices:
self.assertTrue(
conf_matrix.percent_correct>=70,
msg="We must have trained on each one more or " \
"less correctly. Got %f%% correct on %d labels" %
(conf_matrix.percent_correct,
nlabels))
# Since now we have per split and possibly per label -- lets just find
# mean per each feature per label across splits
sensm = FxMapper('samples', lambda x: np.sum(x),
uattrs=['targets']).forward(sens)
sensgm = maxofabs_sample().forward(sensm) # global max of abs of means
assert_equal(sensgm.shape[0], 1)
assert_equal(sensgm.shape[1], ds.nfeatures)
selected = FixedNElementTailSelector(len(ds.a.bogus_features))(
sensgm.samples[0])
if cfg.getboolean('tests', 'labile', default='yes'):
self.assertEqual(
set(selected),
set(ds.a.nonbogus_features),
msg="At the end we should have selected the right features. "
"Chose %s whenever nonbogus are %s" %
(selected, ds.a.nonbogus_features))
# Now test each one per label
# TODO: collect all failures and spit them out at once --
# that would make it easy to see if the sensitivity
# just has incorrect order of labels assigned
for sens1 in sensm:
labels1 = sens1.targets # labels (1) for this sensitivity
lndim = labels1.ndim
label = labels1[0] # current label
# XXX whole lndim comparison should be gone after
# things get fixed and we arrive here with a tuple!
if lndim == 1: # just a single label
self.assertTrue(label in ulabels)
ilabel_all = np.where(ds.fa.nonbogus_targets == label)[0]
# should have just 1 feature for the label
self.assertEqual(len(ilabel_all), 1)
ilabel = ilabel_all[0]
maxsensi = np.argmax(sens1) # index of max sensitivity
self.assertEqual(
maxsensi, ilabel,
"Maximal sensitivity for %s was found in %i whenever"
" original feature was %i for nonbogus features %s" %
(labels1, maxsensi, ilabel, ds.a.nonbogus_features))
elif lndim == 2 and labels1.shape[1] == 2: # pair of labels
# we should have highest (in abs) coefficients in
# those two labels
maxsensi2 = np.argsort(np.abs(sens1))[0][-2:]
ilabel2 = [
np.where(ds.fa.nonbogus_targets == l)[0][0]
for l in label
]
self.assertEqual(
set(maxsensi2), set(ilabel2),
"Maximal sensitivity for %s was found in %s whenever"
" original features were %s for nonbogus features %s" %
(labels1, maxsensi2, ilabel2, ds.a.nonbogus_features))
"""
# Now test for the sign of each one in pair ;) in
# all binary problems L1 (-1) -> L2(+1), then
# weights for L2 should be positive. to test for
# L1 -- invert the sign
# We already know (if we haven't failed in previous test),
# that those 2 were the strongest -- so check only signs
"""
self.assertTrue(
sens1.samples[0, ilabel2[0]] < 0,
"With %i classes in pair %s got feature %i for %r >= 0"
% (nlabels, label, ilabel2[0], label[0]))
self.assertTrue(
sens1.samples[0, ilabel2[1]] > 0,
"With %i classes in pair %s got feature %i for %r <= 0"
% (nlabels, label, ilabel2[1], label[1]))
else:
# yoh could be wrong at this assumption... time will show
self.fail("Got unknown number labels per sensitivity: %s."
" Should be either a single label or a pair" %
labels1)
def test_rfe_sensmap():
# http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/2013q3/002538.html
# just a smoke test. fails with
from mvpa2.clfs.svm import LinearCSVMC
from mvpa2.clfs.meta import FeatureSelectionClassifier
from mvpa2.measures.base import CrossValidation, RepeatedMeasure
from mvpa2.generators.splitters import Splitter
from mvpa2.generators.partition import NFoldPartitioner
from mvpa2.misc.errorfx import mean_mismatch_error
from mvpa2.mappers.fx import mean_sample
from mvpa2.mappers.fx import maxofabs_sample
from mvpa2.generators.base import Repeater
from mvpa2.featsel.rfe import RFE
from mvpa2.featsel.helpers import FractionTailSelector, BestDetector
from mvpa2.featsel.helpers import NBackHistoryStopCrit
from mvpa2.datasets import vstack
from mvpa2.misc.data_generators import normal_feature_dataset
# Let's simulate the beast -- 6 categories total groupped into 3
# super-ordinate, and actually without any 'superordinate' effect
# since subordinate categories independent
fds = normal_feature_dataset(nlabels=3,
snr=1, # 100, # pure signal! ;)
perlabel=9,
nfeatures=6,
nonbogus_features=range(3),
nchunks=3)
clfsvm = LinearCSVMC()
rfesvm = RFE(clfsvm.get_sensitivity_analyzer(postproc=maxofabs_sample()),
CrossValidation(
clfsvm,
NFoldPartitioner(),
errorfx=mean_mismatch_error, postproc=mean_sample()),
Repeater(2),
fselector=FractionTailSelector(0.70, mode='select', tail='upper'),
stopping_criterion=NBackHistoryStopCrit(BestDetector(), 10),
update_sensitivity=True)
fclfsvm = FeatureSelectionClassifier(clfsvm, rfesvm)
sensanasvm = fclfsvm.get_sensitivity_analyzer(postproc=maxofabs_sample())
# manually repeating/splitting so we do both RFE sensitivity and classification
senses, errors = [], []
for i, pset in enumerate(NFoldPartitioner().generate(fds)):
# split partitioned dataset
split = [d for d in Splitter('partitions').generate(pset)]
senses.append(sensanasvm(split[0])) # and it also should train the classifier so we would ask it about error
errors.append(mean_mismatch_error(fclfsvm.predict(split[1]), split[1].targets))
senses = vstack(senses)
errors = vstack(errors)
# Let's compare against rerunning the beast simply for classification with CV
errors_cv = CrossValidation(fclfsvm, NFoldPartitioner(), errorfx=mean_mismatch_error)(fds)
# and they should match
assert_array_equal(errors, errors_cv)
# buggy!
cv_sensana_svm = RepeatedMeasure(sensanasvm, NFoldPartitioner())
senses_rm = cv_sensana_svm(fds)
#print senses.samples, senses_rm.samples
#print errors, errors_cv.samples
assert_raises(AssertionError,
assert_array_almost_equal,
senses.samples, senses_rm.samples)
raise SkipTest("Known failure for repeated measures: https://github.com/PyMVPA/PyMVPA/issues/117")
def test_rfe(self, clf):
# sensitivity analyser and transfer error quantifier use the SAME clf!
sens_ana = clf.get_sensitivity_analyzer(postproc=maxofabs_sample())
pmeasure = ProxyMeasure(clf,
postproc=BinaryFxNode(mean_mismatch_error,
'targets'))
cvmeasure = CrossValidation(clf,
NFoldPartitioner(),
errorfx=mean_mismatch_error,
postproc=mean_sample())
rfesvm_split = SplitClassifier(clf, OddEvenPartitioner())
# explore few recipes
for rfe, data in [
# because the clf is already trained when computing the sensitivity
# map, prevent retraining for transfer error calculation
# Use absolute of the svm weights as sensitivity
(RFE(sens_ana,
pmeasure,
Splitter('train'),
fselector=FixedNElementTailSelector(1),
train_pmeasure=False), self.get_data()),
# use cross-validation within training to get error for the stopping point
# but use full training data to derive sensitivity
(
RFE(
sens_ana,
cvmeasure,
Repeater(
2
), # give the same full dataset to sens_ana and cvmeasure
fselector=FractionTailSelector(0.70,
mode='select',
tail='upper'),
train_pmeasure=True),
normal_feature_dataset(perlabel=20,
nchunks=5,
nfeatures=200,
nonbogus_features=[0, 1],
snr=1.5)),
# use cross-validation (via SplitClassifier) and get mean
# of normed sensitivities across those splits
(
RFE(
rfesvm_split.get_sensitivity_analyzer(
postproc=ChainMapper([
FxMapper('features', l2_normed),
FxMapper('samples', np.mean),
FxMapper('samples', np.abs)
])),
ConfusionBasedError(rfesvm_split, confusion_state='stats'),
Repeater(
2), # we will use the same full cv-training dataset
fselector=FractionTailSelector(0.50,
mode='select',
tail='upper'),
stopping_criterion=NBackHistoryStopCrit(
BestDetector(), 10),
train_pmeasure=
False, # we just extract it from existing confusion
update_sensitivity=True),
normal_feature_dataset(perlabel=28,
nchunks=7,
nfeatures=200,
nonbogus_features=[0, 1],
snr=1.5))
]:
# prep data
# data = datasets['uni2medium']
data_nfeatures = data.nfeatures
rfe.train(data)
resds = rfe(data)
# fail if orig datasets are changed
self.assertTrue(data.nfeatures == data_nfeatures)
# check that the features set with the least error is selected
if len(rfe.ca.errors):
e = np.array(rfe.ca.errors)
if isinstance(rfe._fselector, FixedNElementTailSelector):
self.assertTrue(resds.nfeatures == data_nfeatures -
e.argmin())
else:
imin = np.argmin(e)
if 'does_feature_selection' in clf.__tags__:
# if clf is smart it might figure it out right away
assert_array_less(imin, len(e))
else:
# in this case we can even check if we had actual
# going down/up trend... although -- why up???
self.assertTrue(1 < imin < len(e) - 1)
else:
self.assertTrue(resds.nfeatures == data_nfeatures)
# silly check if nfeatures is in decreasing order
nfeatures = np.array(rfe.ca.nfeatures).copy()
nfeatures.sort()
self.assertTrue((nfeatures[::-1] == rfe.ca.nfeatures).all())
# check if history has elements for every step
self.assertTrue(
set(rfe.ca.history) == set(range(len(np.array(
rfe.ca.errors)))))
# Last (the largest number) can be present multiple times even
# if we remove 1 feature at a time -- just need to stop well
# in advance when we have more than 1 feature left ;)
self.assertTrue(rfe.ca.nfeatures[-1] == len(
np.where(rfe.ca.history == max(rfe.ca.history))[0]))
def __init__(
self,
lrn,
partitioner,
fselector,
errorfx=mean_mismatch_error,
fmeasure_postproc=None,
fmeasure=None,
nproc=1,
# callback?
**kwargs):
"""
Parameters
----------
lrn : Learner
Learner with a sensitivity analyzer which will be used both
for the sensitivity analysis and transfer error estimation
partitioner : Partitioner
Used to generate cross-validation partitions for cross-validation
to deduce optimal number of features to maintain
fselector : Functor
Given a sensitivity map it has to return the ids of those
features that should be kept.
errorfx : func, optional
Functor to use for estimation of cross-validation error
fmeasure_postproc : func, optional
Function to provide to the sensitivity analyzer as postproc. If no
fmeasure is provided and classifier sensitivity is used, then
maxofabs_sample() would be used for this postproc, unless other
value is provided
fmeasure : Function, optional
Featurewise measure. If None was provided, lrn's sensitivity
analyzer will be used.
"""
# Initialize itself preparing for the 2nd invocation
# with determined number of nfeatures_min
# TODO: move this into _train since better not to assign anything here
# to avoid possible problems with copies needing to deal with the same
# lrn... but then we might like again to reconsider delegation instead
# of subclassing here....
if fmeasure is None:
if __debug__:
debug(
'RFE', 'fmeasure was not provided, will be using the '
'sensitivity analyzer for %s' % lrn)
fmeasure = lrn.get_sensitivity_analyzer(
postproc=fmeasure_postproc
if fmeasure_postproc is not None else maxofabs_sample())
train_pmeasure = False
else:
assert fmeasure_postproc is None, "There should be no explicit " \
"fmeasure_postproc when fmeasure is specified"
# if user provided explicit value -- use it! otherwise, we do want
# to train an arbitrary fmeasure
train_pmeasure = kwargs.pop('train_pmeasure', True)
RFE.__init__(self,
fmeasure,
None,
Repeater(2),
fselector=fselector,
bestdetector=None,
train_pmeasure=train_pmeasure,
stopping_criterion=None,
**kwargs)
self._lrn = lrn # should not be modified, thus _
self.partitioner = partitioner
self.errorfx = errorfx
self.fmeasure_postproc = fmeasure_postproc
self.nproc = nproc
def test_rfe(self, clf):
# sensitivity analyser and transfer error quantifier use the SAME clf!
sens_ana = clf.get_sensitivity_analyzer(postproc=maxofabs_sample())
pmeasure = ProxyMeasure(clf, postproc=BinaryFxNode(mean_mismatch_error,
'targets'))
cvmeasure = CrossValidation(clf, NFoldPartitioner(),
errorfx=mean_mismatch_error,
postproc=mean_sample())
rfesvm_split = SplitClassifier(clf, OddEvenPartitioner())
# explore few recipes
for rfe, data in [
# because the clf is already trained when computing the sensitivity
# map, prevent retraining for transfer error calculation
# Use absolute of the svm weights as sensitivity
(RFE(sens_ana,
pmeasure,
Splitter('train'),
fselector=FixedNElementTailSelector(1),
train_pmeasure=False),
self.get_data()),
# use cross-validation within training to get error for the stopping point
# but use full training data to derive sensitivity
(RFE(sens_ana,
cvmeasure,
Repeater(2), # give the same full dataset to sens_ana and cvmeasure
fselector=FractionTailSelector(
0.70,
mode='select', tail='upper'),
train_pmeasure=True),
normal_feature_dataset(perlabel=20, nchunks=5, nfeatures=200,
nonbogus_features=[0, 1], snr=1.5)),
# use cross-validation (via SplitClassifier) and get mean
# of normed sensitivities across those splits
(RFE(rfesvm_split.get_sensitivity_analyzer(
postproc=ChainMapper([ FxMapper('features', l2_normed),
FxMapper('samples', np.mean),
FxMapper('samples', np.abs)])),
ConfusionBasedError(rfesvm_split, confusion_state='stats'),
Repeater(2), # we will use the same full cv-training dataset
fselector=FractionTailSelector(
0.50,
mode='select', tail='upper'),
stopping_criterion=NBackHistoryStopCrit(BestDetector(), 10),
train_pmeasure=False, # we just extract it from existing confusion
update_sensitivity=True),
normal_feature_dataset(perlabel=28, nchunks=7, nfeatures=200,
nonbogus_features=[0, 1], snr=1.5))
]:
# prep data
# data = datasets['uni2medium']
data_nfeatures = data.nfeatures
rfe.train(data)
resds = rfe(data)
# fail if orig datasets are changed
self.assertTrue(data.nfeatures == data_nfeatures)
# check that the features set with the least error is selected
if len(rfe.ca.errors):
e = np.array(rfe.ca.errors)
if isinstance(rfe._fselector, FixedNElementTailSelector):
self.assertTrue(resds.nfeatures == data_nfeatures - e.argmin())
else:
imin = np.argmin(e)
if 'does_feature_selection' in clf.__tags__:
# if clf is smart it might figure it out right away
assert_array_less( imin, len(e) )
else:
# in this case we can even check if we had actual
# going down/up trend... although -- why up???
self.assertTrue( 1 < imin < len(e) - 1 )
else:
self.assertTrue(resds.nfeatures == data_nfeatures)
# silly check if nfeatures is in decreasing order
nfeatures = np.array(rfe.ca.nfeatures).copy()
nfeatures.sort()
self.assertTrue( (nfeatures[::-1] == rfe.ca.nfeatures).all() )
# check if history has elements for every step
self.assertTrue(set(rfe.ca.history)
== set(range(len(np.array(rfe.ca.errors)))))
# Last (the largest number) can be present multiple times even
# if we remove 1 feature at a time -- just need to stop well
# in advance when we have more than 1 feature left ;)
self.assertTrue(rfe.ca.nfeatures[-1]
== len(np.where(rfe.ca.history
==max(rfe.ca.history))[0]))
def test_analyzer_with_split_classifier(self, clfds):
"""Test analyzers in split classifier
"""
clf, ds = clfds # unroll the tuple
# We need to skip some LARSes here
_sclf = str(clf)
if 'LARS(' in _sclf and "type='stepwise'" in _sclf:
# ADD KnownToFail thingie from NiPy
return
# To don't waste too much time testing lets limit to 3 splits
nsplits = 3
partitioner = NFoldPartitioner(count=nsplits)
mclf = SplitClassifier(clf=clf,
partitioner=partitioner,
enable_ca=['training_stats',
'stats'])
sana = mclf.get_sensitivity_analyzer(# postproc=absolute_features(),
pass_attr=['fa.nonbogus_targets'],
enable_ca=["sensitivities"])
ulabels = ds.uniquetargets
nlabels = len(ulabels)
# Can't rely on splitcfg since count-limit is done in __call__
assert(nsplits == len(list(partitioner.generate(ds))))
sens = sana(ds)
assert('nonbogus_targets' in sens.fa) # were they passsed?
# TODO: those few do not expose biases
if not len(set(clf.__tags__).intersection(('lars', 'glmnet', 'gpr'))):
assert('biases' in sens.sa)
# print sens.sa.biases
# It should return either ...
# nlabels * nsplits
req_nsamples = [ nlabels * nsplits ]
if nlabels == 2:
# A single sensitivity in case of binary
req_nsamples += [ nsplits ]
else:
# and for pairs in case of multiclass
req_nsamples += [ (nlabels * (nlabels - 1) / 2) * nsplits ]
# and for 1-vs-1 embedded within Multiclass operating on
# pairs (e.g. SMLR)
req_nsamples += [req_nsamples[-1] * 2]
# Also for regression_based -- they can do multiclass
# but only 1 sensitivity is provided
if 'regression_based' in clf.__tags__:
req_nsamples += [ nsplits ]
# # of features should correspond
self.assertEqual(sens.shape[1], ds.nfeatures)
# # of samples/sensitivities should also be reasonable
self.assertTrue(sens.shape[0] in req_nsamples)
# Check if labels are present
self.assertTrue('splits' in sens.sa)
self.assertTrue('targets' in sens.sa)
# should be 1D -- otherwise dtype object
self.assertTrue(sens.sa.targets.ndim == 1)
sens_ulabels = sens.sa['targets'].unique
# Some labels might be pairs(tuples) so ndarray would be of
# dtype object and we would need to get them all
if sens_ulabels.dtype is np.dtype('object'):
sens_ulabels = np.unique(
reduce(lambda x, y: x + y, [list(x) for x in sens_ulabels]))
assert_array_equal(sens_ulabels, ds.sa['targets'].unique)
errors = [x.percent_correct
for x in sana.clf.ca.stats.matrices]
# lets go through all sensitivities and see if we selected the right
# features
#if 'meta' in clf.__tags__ and len(sens.samples[0].nonzero()[0])<2:
if '5%' in clf.descr \
or (nlabels > 2 and 'regression_based' in clf.__tags__):
# Some meta classifiers (5% of ANOVA) are too harsh ;-)
# if we get less than 2 features with on-zero sensitivities we
# cannot really test
# Also -- regression based classifiers performance for multiclass
# is expected to suck in general
return
if cfg.getboolean('tests', 'labile', default='yes'):
for conf_matrix in [sana.clf.ca.training_stats] \
+ sana.clf.ca.stats.matrices:
self.assertTrue(
conf_matrix.percent_correct >= 70,
msg="We must have trained on each one more or " \
"less correctly. Got %f%% correct on %d labels" %
(conf_matrix.percent_correct,
nlabels))
# Since now we have per split and possibly per label -- lets just find
# mean per each feature per label across splits
sensm = FxMapper('samples', lambda x: np.sum(x),
uattrs=['targets']).forward(sens)
sensgm = maxofabs_sample().forward(sensm) # global max of abs of means
assert_equal(sensgm.shape[0], 1)
assert_equal(sensgm.shape[1], ds.nfeatures)
selected = FixedNElementTailSelector(
len(ds.a.bogus_features))(sensgm.samples[0])
if cfg.getboolean('tests', 'labile', default='yes'):
self.assertEqual(
set(selected), set(ds.a.nonbogus_features),
msg="At the end we should have selected the right features. "
"Chose %s whenever nonbogus are %s"
% (selected, ds.a.nonbogus_features))
# Now test each one per label
# TODO: collect all failures and spit them out at once --
# that would make it easy to see if the sensitivity
# just has incorrect order of labels assigned
for sens1 in sensm:
labels1 = sens1.targets # labels (1) for this sensitivity
lndim = labels1.ndim
label = labels1[0] # current label
# XXX whole lndim comparison should be gone after
# things get fixed and we arrive here with a tuple!
if lndim == 1: # just a single label
self.assertTrue(label in ulabels)
ilabel_all = np.where(ds.fa.nonbogus_targets == label)[0]
# should have just 1 feature for the label
self.assertEqual(len(ilabel_all), 1)
ilabel = ilabel_all[0]
maxsensi = np.argmax(sens1) # index of max sensitivity
self.assertEqual(maxsensi, ilabel,
"Maximal sensitivity for %s was found in %i whenever"
" original feature was %i for nonbogus features %s"
% (labels1, maxsensi, ilabel, ds.a.nonbogus_features))
elif lndim == 2 and labels1.shape[1] == 2: # pair of labels
# we should have highest (in abs) coefficients in
# those two labels
maxsensi2 = np.argsort(np.abs(sens1))[0][-2:]
ilabel2 = [np.where(ds.fa.nonbogus_targets == l)[0][0]
for l in label]
self.assertEqual(
set(maxsensi2), set(ilabel2),
"Maximal sensitivity for %s was found in %s whenever"
" original features were %s for nonbogus features %s"
% (labels1, maxsensi2, ilabel2, ds.a.nonbogus_features))
"""
# Now test for the sign of each one in pair ;) in
# all binary problems L1 (-1) -> L2(+1), then
# weights for L2 should be positive. to test for
# L1 -- invert the sign
# We already know (if we haven't failed in previous test),
# that those 2 were the strongest -- so check only signs
"""
self.assertTrue(
sens1.samples[0, ilabel2[0]] < 0,
"With %i classes in pair %s got feature %i for %r >= 0"
% (nlabels, label, ilabel2[0], label[0]))
self.assertTrue(sens1.samples[0, ilabel2[1]] > 0,
"With %i classes in pair %s got feature %i for %r <= 0"
% (nlabels, label, ilabel2[1], label[1]))
else:
# yoh could be wrong at this assumption... time will show
self.fail("Got unknown number labels per sensitivity: %s."
" Should be either a single label or a pair"
% labels1)
def __repr__(self, prefixes=[]):
return super(SplitRFE, self).__repr__(
prefixes=prefixes + _repr_attrs(self, ['lrn', 'partitioner']) +
_repr_attrs(self, ['errorfx'], default=mean_mismatch_error) +
_repr_attrs(self, ['analyzer_postproc'], default=maxofabs_sample())
)
descr='skl.LassoLarsIC()')
regrswh += [_lasso_lars_ic]
clfswh += [
RegressionAsClassifier(_lasso_lars_ic, descr='skl.LassoLarsIC_C()')
]
# kNN
clfswh += kNN(k=5, descr="kNN(k=5)")
clfswh += kNN(k=5, voting='majority', descr="kNN(k=5, voting='majority')")
clfswh += \
FeatureSelectionClassifier(
kNN(),
SensitivityBasedFeatureSelection(
SMLRWeights(SMLR(lm=1.0, implementation="C"),
postproc=maxofabs_sample()),
RangeElementSelector(mode='select')),
descr="kNN on SMLR(lm=1) non-0")
clfswh += \
FeatureSelectionClassifier(
kNN(),
SensitivityBasedFeatureSelection(
OneWayAnova(),
FractionTailSelector(0.05, mode='select', tail='upper')),
descr="kNN on 5%(ANOVA)")
clfswh += \
FeatureSelectionClassifier(
kNN(),
SensitivityBasedFeatureSelection(
tags=_lars_tags,
descr='skl.LassoLarsIC()')
regrswh += [_lasso_lars_ic]
clfswh += [RegressionAsClassifier(_lasso_lars_ic,
descr='skl.LassoLarsIC_C()')]
# kNN
clfswh += kNN(k=5, descr="kNN(k=5)")
clfswh += kNN(k=5, voting='majority', descr="kNN(k=5, voting='majority')")
clfswh += \
FeatureSelectionClassifier(
kNN(),
SensitivityBasedFeatureSelection(
SMLRWeights(SMLR(lm=1.0, implementation="C"),
postproc=maxofabs_sample()),
RangeElementSelector(mode='select')),
descr="kNN on SMLR(lm=1) non-0")
clfswh += \
FeatureSelectionClassifier(
kNN(),
SensitivityBasedFeatureSelection(
OneWayAnova(),
FractionTailSelector(0.05, mode='select', tail='upper')),
descr="kNN on 5%(ANOVA)")
clfswh += \
FeatureSelectionClassifier(
kNN(),
SensitivityBasedFeatureSelection(
_lasso_lars = SKLLearnerAdapter(sklLassoLARS(), tags=_lars_tags, descr="skl.LassoLARS()")
regrswh += [_lars, _lasso_lars]
clfswh += [
RegressionAsClassifier(_lars, descr="skl.LARS_C()"),
RegressionAsClassifier(_lasso_lars, descr="skl.LassoLARS_C()"),
]
# kNN
clfswh += kNN(k=5, descr="kNN(k=5)")
clfswh += kNN(k=5, voting="majority", descr="kNN(k=5, voting='majority')")
clfswh += FeatureSelectionClassifier(
kNN(),
SensitivityBasedFeatureSelection(
SMLRWeights(SMLR(lm=1.0, implementation="C"), postproc=maxofabs_sample()), RangeElementSelector(mode="select")
),
descr="kNN on SMLR(lm=1) non-0",
)
clfswh += FeatureSelectionClassifier(
kNN(),
SensitivityBasedFeatureSelection(OneWayAnova(), FractionTailSelector(0.05, mode="select", tail="upper")),
descr="kNN on 5%(ANOVA)",
)
clfswh += FeatureSelectionClassifier(
kNN(),
SensitivityBasedFeatureSelection(OneWayAnova(), FixedNElementTailSelector(50, mode="select", tail="upper")),
descr="kNN on 50(ANOVA)",
)
