def test_diff_len_labels_str_treeclassifier(self):
# check if the classifier can handle a dataset with labels as string of
# variable length
# was failing on TreeClassifier due to np.str dtype being assumed from first
# returned value
ds = datasets['uni4small'].copy()
newlabels = dict([(l,l+'_'*li) for li,l in enumerate(ds.uniquetargets)])
ds.targets = [newlabels[l] for l in ds.targets]
clf = TreeClassifier(mvpa2.testing.clfs.SVM(), {
'group1':(ds.uniquetargets[:2], mvpa2.testing.clfs.SVM()),
'group2':(ds.uniquetargets[2:], mvpa2.testing.clfs.SVM())})
clf.train(ds)
predictions = clf.predict(ds)
# predictions on the same ds as training should give same labels
assert(np.all(np.unique(predictions) == ds.uniquetargets))
def test_diff_len_labels_str_treeclassifier(self):
# check if the classifier can handle a dataset with labels as string of
# variable length
# was failing on TreeClassifier due to np.str dtype being assumed from first
# returned value
ds = datasets['uni4small'].copy()
newlabels = dict([(l, l + '_' * li)
for li, l in enumerate(ds.uniquetargets)])
ds.targets = [newlabels[l] for l in ds.targets]
clf = TreeClassifier(
mvpa2.testing.clfs.SVM(), {
'group1': (ds.uniquetargets[:2], mvpa2.testing.clfs.SVM()),
'group2': (ds.uniquetargets[2:], mvpa2.testing.clfs.SVM())
})
clf.train(ds)
predictions = clf.predict(ds)
# predictions on the same ds as training should give same labels
assert (np.all(np.unique(predictions) == ds.uniquetargets))
def test_tree_classifier(self):
"""Basic tests for TreeClassifier
"""
ds = datasets['uni4medium']
# make it simple of the beast -- take only informative ones
# because classifiers for the tree are selected randomly, so
# performance varies a lot and we just need to check on
# correct operation
ds = ds[:, ds.fa.nonbogus_targets != [None]]
clfs = clfswh['binary'] # pool of classifiers
# Lets permute so each time we try some different combination
# of the classifiers but exclude those operating on %s of
# features since we might not have enough for that
clfs = [
clfs[i] for i in np.random.permutation(len(clfs))
if not '%' in str(clfs[i])
]
# NB: It is necessary that the same classifier was not used at
# different nodes, since it would be re-trained for a new set
# of targets, thus leading to incorrect behavior/high error.
#
# Clone only those few leading ones which we will use
# throughout the test
clfs = [clf.clone() for clf in clfs[:4]]
# Test conflicting definition
tclf = TreeClassifier(clfs[0], {
'L0+2': (('L0', 'L2'), clfs[1]),
'L2+3': (('L2', 'L3'), clfs[2])
})
self.assertRaises(ValueError, tclf.train, ds)
"""Should raise exception since label 2 is in both"""
# Test insufficient definition
tclf = TreeClassifier(clfs[0], {
'L0+5': (('L0', 'L5'), clfs[1]),
'L2+3': (('L2', 'L3'), clfs[2])
})
self.assertRaises(ValueError, tclf.train, ds)
"""Should raise exception since no group for L1"""
# proper definition now
tclf = TreeClassifier(clfs[0], {
'L0+1': (('L0', 'L1'), clfs[1]),
'L2+3': (('L2', 'L3'), clfs[2])
})
# Lets test train/test cycle using CVTE
cv = CrossValidation(tclf,
OddEvenPartitioner(),
postproc=mean_sample(),
enable_ca=['stats', 'training_stats'])
cverror = cv(ds).samples.squeeze()
try:
rtclf = repr(tclf)
except:
self.fail(msg="Could not obtain repr for TreeClassifier")
# Test accessibility of .clfs
self.assertTrue(tclf.clfs['L0+1'] is clfs[1])
self.assertTrue(tclf.clfs['L2+3'] is clfs[2])
cvtrc = cv.ca.training_stats
cvtc = cv.ca.stats
if cfg.getboolean('tests', 'labile', default='yes'):
# just a dummy check to make sure everything is working
self.assertTrue(cvtrc != cvtc)
self.assertTrue(cverror < 0.3,
msg="Got too high error = %s using %s" %
(cverror, tclf))
# Test trailing nodes with no classifier
# That is why we use separate pool of classifiers here
# (that is probably old/not-needed since switched to use clones)
clfs_mc = clfswh['multiclass'] # pool of classifiers
clfs_mc = [
clfs_mc[i] for i in np.random.permutation(len(clfs_mc))
if not '%' in str(clfs_mc[i])
]
clfs_mc = [clf.clone() for clf in clfs_mc[:4]] # and clones again
tclf = TreeClassifier(clfs_mc[0], {
'L0': (('L0', ), None),
'L1+2+3': (('L1', 'L2', 'L3'), clfs_mc[1])
})
cv = CrossValidation(tclf,
OddEvenPartitioner(),
postproc=mean_sample(),
enable_ca=['stats', 'training_stats'])
cverror = np.asscalar(cv(ds))
if cfg.getboolean('tests', 'labile', default='yes'):
self.assertTrue(cverror < 0.3,
msg="Got too high error = %s using %s" %
(cverror, tclf))