def test_feature_selection_pipeline(self):
sens_ana = SillySensitivityAnalyzer()
data = self.get_data()
data_nfeatures = data.nfeatures
# test silly one first ;-)
self.assertEqual(sens_ana(data).samples[0,0], -int(data_nfeatures/2))
# OLD: first remove 25% == 6, and then 4, total removing 10
# NOW: test should be independent of the numerical number of features
feature_selections = [SensitivityBasedFeatureSelection(
sens_ana,
FractionTailSelector(0.25)),
SensitivityBasedFeatureSelection(
sens_ana,
FixedNElementTailSelector(4))
]
# create a FeatureSelection pipeline
feat_sel_pipeline = ChainMapper(feature_selections)
feat_sel_pipeline.train(data)
resds = feat_sel_pipeline(data)
self.assertEqual(len(feat_sel_pipeline),
len(feature_selections),
msg="Test the property feature_selections")
desired_nfeatures = int(np.ceil(data_nfeatures*0.75))
self.assertEqual([fe._oshape[0] for fe in feat_sel_pipeline],
[desired_nfeatures, desired_nfeatures - 4])
def test_chainmapper():
# the chain needs at lest one mapper
assert_raises(ValueError, ChainMapper, [])
# a typical first mapper is to flatten
cm = ChainMapper([FlattenMapper()])
# few container checks
assert_equal(len(cm), 1)
assert_true(isinstance(cm[0], FlattenMapper))
# now training
# come up with data
samples_shape = (2, 2, 4)
data_shape = (4,) + samples_shape
data = np.arange(np.prod(data_shape)).reshape(data_shape)
pristinedata = data.copy()
target = [[ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15],
[16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31],
[32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47],
[48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63]]
target = np.array(target)
# if it is not trained it knows nothing
cm.train(data)
# a new mapper should appear when doing feature selection
cm.append(StaticFeatureSelection(range(1, 16)))
assert_equal(cm.forward1(data[0]).shape, (15,))
assert_equal(len(cm), 2)
# multiple slicing
cm.append(StaticFeatureSelection([9, 14]))
assert_equal(cm.forward1(data[0]).shape, (2,))
assert_equal(len(cm), 3)
# check reproduction
if __debug__:
# debug mode needs special test as it enhances the repr output
# with module info and id() appendix for objects
import mvpa2
cm_clone = eval(repr(cm))
assert_equal('#'.join(repr(cm_clone).split('#')[:-1]),
'#'.join(repr(cm).split('#')[:-1]))
else:
cm_clone = eval(repr(cm))
assert_equal(repr(cm_clone), repr(cm))
# what happens if we retrain the whole beast an same data as before
cm.train(data)
assert_equal(cm.forward1(data[0]).shape, (2,))
assert_equal(len(cm), 3)
# let's map something
mdata = cm.forward(data)
assert_array_equal(mdata, target[:, [10, 15]])
# and back
rdata = cm.reverse(mdata)
# original shape
assert_equal(rdata.shape, data.shape)
# content as far it could be restored
assert_array_equal(rdata[rdata > 0], data[rdata > 0])
assert_equal(np.sum(rdata > 0), 8)
# Lets construct a dataset with mapper assigned and see
# if sub-selecting a feature adjusts trailing StaticFeatureSelection
# appropriately
ds_subsel = Dataset.from_wizard(data, mapper=cm)[:, 1]
tail_sfs = ds_subsel.a.mapper[-1]
assert_equal(repr(tail_sfs), 'StaticFeatureSelection(slicearg=array([14]))')
def test_datasetmapping():
# 6 samples, 4X2 features
data = np.arange(48).reshape(6, 4, 2)
ds = Dataset(data,
sa={
'timepoints': np.arange(6),
'multidim': data.copy()
},
fa={'fid': np.arange(4)})
# with overlapping and non-overlapping boxcars
startpoints = [0, 1, 4]
boxlength = 2
bm = BoxcarMapper(startpoints, boxlength, space='boxy')
# train is critical
bm.train(ds)
mds = bm.forward(ds)
assert_equal(len(mds), len(startpoints))
assert_equal(mds.nfeatures, boxlength)
# all samples attributes remain, but the can rotated/compressed into
# multidimensional attributes
assert_equal(sorted(mds.sa.keys()),
['boxy_onsetidx'] + sorted(ds.sa.keys()))
assert_equal(mds.sa.multidim.shape,
(len(startpoints), boxlength) + ds.shape[1:])
assert_equal(mds.sa.timepoints.shape, (len(startpoints), boxlength))
assert_array_equal(mds.sa.timepoints.flatten(),
np.array([(s, s + 1) for s in startpoints]).flatten())
assert_array_equal(mds.sa.boxy_onsetidx, startpoints)
# feature attributes also get rotated and broadcasted
assert_array_equal(mds.fa.fid, [ds.fa.fid, ds.fa.fid])
# and finally there is a new one
assert_array_equal(mds.fa.boxy_offsetidx, range(boxlength))
# now see how it works on reverse()
rds = bm.reverse(mds)
# we got at least something of all original attributes back
assert_equal(sorted(rds.sa.keys()), sorted(ds.sa.keys()))
assert_equal(sorted(rds.fa.keys()), sorted(ds.fa.keys()))
# it is not possible to reconstruct the full samples array
# some samples even might show up multiple times (when there are overlapping
# boxcars
assert_array_equal(
rds.samples,
np.array([[[0, 1], [2, 3], [4, 5], [6, 7]],
[[8, 9], [10, 11], [12, 13], [14, 15]],
[[8, 9], [10, 11], [12, 13], [14, 15]],
[[16, 17], [18, 19], [20, 21], [22, 23]],
[[32, 33], [34, 35], [36, 37], [38, 39]],
[[40, 41], [42, 43], [44, 45], [46, 47]]]))
assert_array_equal(rds.sa.timepoints, [0, 1, 1, 2, 4, 5])
assert_array_equal(rds.sa.multidim, ds.sa.multidim[rds.sa.timepoints])
# but feature attributes should be fully recovered
assert_array_equal(rds.fa.fid, ds.fa.fid)
# popular dataset configuration (double flatten + boxcar)
cm = ChainMapper([FlattenMapper(), bm, FlattenMapper()])
cm.train(ds)
bflat = ds.get_mapped(cm)
assert_equal(bflat.shape,
(len(startpoints), boxlength * np.prod(ds.shape[1:])))
# add attributes
bflat.fa['testfa'] = np.arange(bflat.nfeatures)
bflat.sa['testsa'] = np.arange(bflat.nsamples)
# now try to go back
bflatrev = bflat.mapper.reverse(bflat)
# data should be same again, as far as the boxcars match
assert_array_equal(ds.samples[:2], bflatrev.samples[:2])
assert_array_equal(ds.samples[-2:], bflatrev.samples[-2:])
# feature axis should match
assert_equal(ds.shape[1:], bflatrev.shape[1:])
def test_datasetmapping():
# 6 samples, 4X2 features
data = np.arange(48).reshape(6,4,2)
ds = Dataset(data,
sa={'timepoints': np.arange(6),
'multidim': data.copy()},
fa={'fid': np.arange(4)})
# with overlapping and non-overlapping boxcars
startpoints = [0, 1, 4]
boxlength = 2
bm = BoxcarMapper(startpoints, boxlength, space='boxy')
# train is critical
bm.train(ds)
mds = bm.forward(ds)
assert_equal(len(mds), len(startpoints))
assert_equal(mds.nfeatures, boxlength)
# all samples attributes remain, but the can rotated/compressed into
# multidimensional attributes
assert_equal(sorted(mds.sa.keys()), ['boxy_onsetidx'] + sorted(ds.sa.keys()))
assert_equal(mds.sa.multidim.shape,
(len(startpoints), boxlength) + ds.shape[1:])
assert_equal(mds.sa.timepoints.shape, (len(startpoints), boxlength))
assert_array_equal(mds.sa.timepoints.flatten(),
np.array([(s, s+1) for s in startpoints]).flatten())
assert_array_equal(mds.sa.boxy_onsetidx, startpoints)
# feature attributes also get rotated and broadcasted
assert_array_equal(mds.fa.fid, [ds.fa.fid, ds.fa.fid])
# and finally there is a new one
assert_array_equal(mds.fa.boxy_offsetidx, range(boxlength))
# now see how it works on reverse()
rds = bm.reverse(mds)
# we got at least something of all original attributes back
assert_equal(sorted(rds.sa.keys()), sorted(ds.sa.keys()))
assert_equal(sorted(rds.fa.keys()), sorted(ds.fa.keys()))
# it is not possible to reconstruct the full samples array
# some samples even might show up multiple times (when there are overlapping
# boxcars
assert_array_equal(rds.samples,
np.array([[[ 0, 1], [ 2, 3], [ 4, 5], [ 6, 7]],
[[ 8, 9], [10, 11], [12, 13], [14, 15]],
[[ 8, 9], [10, 11], [12, 13], [14, 15]],
[[16, 17], [18, 19], [20, 21], [22, 23]],
[[32, 33], [34, 35], [36, 37], [38, 39]],
[[40, 41], [42, 43], [44, 45], [46, 47]]]))
assert_array_equal(rds.sa.timepoints, [0, 1, 1, 2, 4, 5])
assert_array_equal(rds.sa.multidim, ds.sa.multidim[rds.sa.timepoints])
# but feature attributes should be fully recovered
assert_array_equal(rds.fa.fid, ds.fa.fid)
# popular dataset configuration (double flatten + boxcar)
cm= ChainMapper([FlattenMapper(), bm, FlattenMapper()])
cm.train(ds)
bflat = ds.get_mapped(cm)
assert_equal(bflat.shape, (len(startpoints), boxlength * np.prod(ds.shape[1:])))
# add attributes
bflat.fa['testfa'] = np.arange(bflat.nfeatures)
bflat.sa['testsa'] = np.arange(bflat.nsamples)
# now try to go back
bflatrev = bflat.mapper.reverse(bflat)
# data should be same again, as far as the boxcars match
assert_array_equal(ds.samples[:2], bflatrev.samples[:2])
assert_array_equal(ds.samples[-2:], bflatrev.samples[-2:])
# feature axis should match
assert_equal(ds.shape[1:], bflatrev.shape[1:])
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]))