def test_split_clf_on_chainpartitioner(self):
# pretty much a smoke test for #156
ds = datasets['uni2small']
part = ChainNode([
NFoldPartitioner(cvtype=1),
Balancer(attr='targets',
count=2,
limit='partitions',
apply_selection=True)
])
partitions = list(part.generate(ds))
sclf = SplitClassifier(sample_clf_lin,
part,
enable_ca=['stats', 'splits'])
sclf.train(ds)
pred = sclf.predict(ds)
assert_equal(len(pred), len(ds)) # rudimentary check
assert_equal(len(sclf.ca.splits), len(partitions))
assert_equal(len(sclf.clfs), len(partitions))
# now let's do sensitivity analyzer just in case
sclf.untrain()
sensana = sclf.get_sensitivity_analyzer()
sens = sensana(ds)
# basic check that sensitivities varied across splits
from mvpa2.mappers.fx import FxMapper
sens_stds = FxMapper('samples', np.std, uattrs=['targets'])(sens)
assert_true(np.any(sens_stds != 0))
def test_sifter_with_balancing():
# extended previous test which was already
# "... somewhat duplicating the doctest"
ds = Dataset(samples=np.arange(12).reshape((-1, 2)),
sa={'chunks': [ 0 , 1 , 2 , 3 , 4, 5 ],
'targets': ['c', 'c', 'c', 'p', 'p', 'p']})
# Without sifter -- just to assure that we do get all of them
# i.e. 6*5*4*3/(4!) = 15
par = ChainNode([NFoldPartitioner(cvtype=4, attr='chunks')])
assert_equal(len(list(par.generate(ds))), 15)
# so we will take 4 chunks out of available 7, but would care only
# about those partitions where we have balanced number of 'c' and 'p'
# entries
assert_raises(ValueError,
lambda x: list(Sifter([('targets', dict(wrong=1))]).generate(x)),
ds)
par = ChainNode([NFoldPartitioner(cvtype=4, attr='chunks'),
Sifter([('partitions', 2),
('targets',
dict(uvalues=['c', 'p'],
balanced=True))])
])
dss = list(par.generate(ds))
# print [ x[x.sa.partitions==2].sa.targets for x in dss ]
assert_equal(len(dss), 9)
for ds_ in dss:
testing = ds[ds_.sa.partitions == 2]
assert_array_equal(np.unique(testing.sa.targets), ['c', 'p'])
# and we still have both targets present in training
training = ds[ds_.sa.partitions == 1]
assert_array_equal(np.unique(training.sa.targets), ['c', 'p'])
def test_exclude_targets_combinations_subjectchunks():
partitioner = ChainNode([NFoldPartitioner(attr='subjects'),
ExcludeTargetsCombinationsPartitioner(
k=1,
targets_attr='chunks',
space='partitions')],
space='partitions')
# targets do not need even to be defined!
ds = Dataset(np.arange(18).reshape(9, 2),
sa={'chunks': np.arange(9) // 3,
'subjects': np.arange(9) % 3})
dss = list(partitioner.generate(ds))
assert_equal(len(dss), 9)
testing_subjs, testing_chunks = [], []
for ds_ in dss:
testing_partition = ds_.sa.partitions == 2
training_partition = ds_.sa.partitions == 1
# must be scalars -- so implicit test here
# if not -- would be error
testing_subj = np.asscalar(np.unique(ds_.sa.subjects[testing_partition]))
testing_subjs.append(testing_subj)
testing_chunk = np.asscalar(np.unique(ds_.sa.chunks[testing_partition]))
testing_chunks.append(testing_chunk)
# and those must not appear for training
ok_(not testing_subj in ds_.sa.subjects[training_partition])
ok_(not testing_chunk in ds_.sa.chunks[training_partition])
# and we should have gone through all chunks/subjs pairs
testing_pairs = set(zip(testing_subjs, testing_chunks))
assert_equal(len(testing_pairs), 9)
# yoh: equivalent to set(itertools.product(range(3), range(3))))
# but .product is N/A for python2.5
assert_equal(testing_pairs, set(zip(*np.where(np.ones((3,3))))))
def _test_edmund_chong_20120907(): # pragma: no cover
# commented out to avoid syntax warnings while compiling
# from mvpa2.suite import *
from mvpa2.testing.datasets import datasets
repeater = Repeater(count=20)
partitioner = ChainNode([NFoldPartitioner(cvtype=1),
Balancer(attr='targets',
count=1, # for real data > 1
limit='partitions',
apply_selection=True
)],
space='partitions')
clf = LinearCSVMC() #choice of classifier
permutator = AttributePermutator('targets', limit={'partitions': 1},
count=1)
null_cv = CrossValidation(
clf,
ChainNode([partitioner, permutator], space=partitioner.get_space()),
errorfx=mean_mismatch_error)
distr_est = MCNullDist(repeater, tail='left', measure=null_cv,
enable_ca=['dist_samples'])
cvte = CrossValidation(clf, partitioner,
errorfx=mean_mismatch_error,
null_dist=distr_est,
enable_ca=['stats'])
errors = cvte(datasets['uni2small'])
def test_permute_superord():
from mvpa2.base.node import ChainNode
from mvpa2.generators.partition import NFoldPartitioner
from mvpa2.generators.base import Sifter
from mvpa2.generators.permutation import AttributePermutator
ds = _get_superord_dataset()
# mvpa2.seed(1)
part = ChainNode([
## so we split based on superord
NFoldPartitioner(len(ds.sa['superord'].unique),
attr='subord'),
## so it should select only those splits where we took 1 from
## each of the superord categories leaving things in balance
Sifter([('partitions', 2),
('superord',
{ 'uvalues': ds.sa['superord'].unique,
'balanced': True})]),
AttributePermutator(['superord'], limit=['partitions',
'chunks']),
], space='partitions')
for ds_perm in part.generate(ds):
# it does permutation
assert(np.sum(ds_perm.sa.superord != ds.sa.superord) != 0)
def test_permute_superord():
from mvpa2.base.node import ChainNode
from mvpa2.generators.partition import NFoldPartitioner
from mvpa2.generators.base import Sifter
from mvpa2.generators.permutation import AttributePermutator
ds = _get_superord_dataset()
# mvpa2.seed(1)
part = ChainNode(
[
## so we split based on superord
NFoldPartitioner(len(ds.sa['superord'].unique), attr='subord'),
## so it should select only those splits where we took 1 from
## each of the superord categories leaving things in balance
Sifter([('partitions', 2),
('superord', {
'uvalues': ds.sa['superord'].unique,
'balanced': True
})]),
AttributePermutator(['superord'], limit=['partitions', 'chunks']),
],
space='partitions')
for ds_perm in part.generate(ds):
# it does permutation
assert (np.sum(ds_perm.sa.superord != ds.sa.superord) != 0)
def test_exclude_targets_combinations_subjectchunks():
partitioner = ChainNode([NFoldPartitioner(attr='subjects'),
ExcludeTargetsCombinationsPartitioner(
k=1,
targets_attr='chunks',
space='partitions')],
space='partitions')
# targets do not need even to be defined!
ds = Dataset(np.arange(18).reshape(9, 2),
sa={'chunks': np.arange(9) // 3,
'subjects': np.arange(9) % 3})
dss = list(partitioner.generate(ds))
assert_equal(len(dss), 9)
testing_subjs, testing_chunks = [], []
for ds_ in dss:
testing_partition = ds_.sa.partitions == 2
training_partition = ds_.sa.partitions == 1
# must be scalars -- so implicit test here
# if not -- would be error
testing_subj = np.asscalar(np.unique(ds_.sa.subjects[testing_partition]))
testing_subjs.append(testing_subj)
testing_chunk = np.asscalar(np.unique(ds_.sa.chunks[testing_partition]))
testing_chunks.append(testing_chunk)
# and those must not appear for training
ok_(not testing_subj in ds_.sa.subjects[training_partition])
ok_(not testing_chunk in ds_.sa.chunks[training_partition])
# and we should have gone through all chunks/subjs pairs
testing_pairs = set(zip(testing_subjs, testing_chunks))
assert_equal(len(testing_pairs), 9)
# yoh: equivalent to set(itertools.product(range(3), range(3))))
# but .product is N/A for python2.5
assert_equal(testing_pairs, set(zip(*np.where(np.ones((3,3))))))
def test_exclude_targets_combinations():
partitioner = ChainNode([
NFoldPartitioner(),
ExcludeTargetsCombinationsPartitioner(
k=2, targets_attr='targets', space='partitions')
],
space='partitions')
from mvpa2.misc.data_generators import normal_feature_dataset
ds = normal_feature_dataset(snr=0.,
nlabels=4,
perlabel=3,
nchunks=3,
nonbogus_features=[0, 1, 2, 3],
nfeatures=4)
partitions = list(partitioner.generate(ds))
assert_equal(len(partitions), 3 * 6)
splitter = Splitter('partitions')
combs = []
comb_chunks = []
for p in partitions:
trds, teds = list(splitter.generate(p))[:2]
comb = tuple(np.unique(teds.targets))
combs.append(comb)
comb_chunks.append(comb + tuple(np.unique(teds.chunks)))
assert_equal(len(set(combs)),
6) # just 6 possible combinations of 2 out of 4
assert_equal(len(set(comb_chunks)), 3 * 6) # all unique
def test_splitter():
ds = give_data()
# split with defaults
spl1 = Splitter('chunks')
assert_raises(NotImplementedError, spl1, ds)
splits = list(spl1.generate(ds))
assert_equal(len(splits), len(ds.sa['chunks'].unique))
for split in splits:
# it should have perform basic slicing!
assert_true(split.samples.base is ds.samples)
assert_equal(len(split.sa['chunks'].unique), 1)
assert_true('lastsplit' in split.a)
assert_true(splits[-1].a.lastsplit)
# now again, more customized
spl2 = Splitter('targets',
attr_values=[0, 1, 1, 2, 3, 3, 3],
count=4,
noslicing=True)
splits = list(spl2.generate(ds))
assert_equal(len(splits), 4)
for split in splits:
# it should NOT have perform basic slicing!
assert_false(split.samples.base is ds.samples)
assert_equal(len(split.sa['targets'].unique), 1)
assert_equal(len(split.sa['chunks'].unique), 10)
assert_true(splits[-1].a.lastsplit)
# two should be identical
assert_array_equal(splits[1].samples, splits[2].samples)
# now go wild and split by feature attribute
ds.fa['roi'] = np.repeat([0, 1], 5)
# splitter should auto-detect that this is a feature attribute
spl3 = Splitter('roi')
splits = list(spl3.generate(ds))
assert_equal(len(splits), 2)
for split in splits:
assert_true(split.samples.base is ds.samples)
assert_equal(len(split.fa['roi'].unique), 1)
assert_equal(split.shape, (100, 5))
# and finally test chained splitters
cspl = ChainNode([spl2, spl3, spl1])
splits = list(cspl.generate(ds))
# 4 target splits and 2 roi splits each and 10 chunks each
assert_equal(len(splits), 80)
def test_noise_classification(self):
# get a dataset with a very high SNR
data = get_mv_pattern(10)
# do crossval with default errorfx and 'mean' combiner
cv = CrossValidation(sample_clf_nl, NFoldPartitioner())
# must return a scalar value
result = cv(data)
# must be perfect
self.assertTrue((result.samples < 0.05).all())
# do crossval with permuted regressors
cv = CrossValidation(
sample_clf_nl,
ChainNode(
[NFoldPartitioner(),
AttributePermutator('targets', count=10)],
space='partitions'))
results = cv(data)
# results must not be the same
self.assertTrue(len(np.unique(results.samples)) > 1)
# must be at chance level
pmean = np.array(results).mean()
self.assertTrue(pmean < 0.58 and pmean > 0.42)
def test_split_featurewise_dataset_measure(self):
ds = datasets['uni3small']
sana = RepeatedMeasure(
SMLR(fit_all_weights=True).get_sensitivity_analyzer(),
ChainNode(
[NFoldPartitioner(),
Splitter('partitions', attr_values=[1])]))
sens = sana(ds)
# a sensitivity for each chunk and each label combination
assert_equal(sens.shape, (len(ds.sa['chunks'].unique) *
len(ds.sa['targets'].unique), ds.nfeatures))
# Lets try more complex example with 'boosting'
ds = datasets['uni3medium']
ds.init_origids('samples')
sana = RepeatedMeasure(
SMLR(fit_all_weights=True).get_sensitivity_analyzer(),
Balancer(amount=0.25, count=2, apply_selection=True),
enable_ca=['datasets', 'repetition_results'])
sens = sana(ds)
assert_equal(sens.shape,
(2 * len(ds.sa['targets'].unique), ds.nfeatures))
splits = sana.ca.datasets
self.assertEqual(len(splits), 2)
self.assertTrue(
np.all([s.nsamples == ds.nsamples // 4 for s in splits]))
# should have used different samples
self.assertTrue(np.any([splits[0].sa.origids != splits[1].sa.origids]))
# and should have got different sensitivities
self.assertTrue(np.any(sens[0] != sens[3]))
def test_splitter():
ds = give_data()
# split with defaults
spl1 = Splitter('chunks')
assert_raises(NotImplementedError, spl1, ds)
splits = list(spl1.generate(ds))
assert_equal(len(splits), len(ds.sa['chunks'].unique))
for split in splits:
# it should have perform basic slicing!
assert_true(split.samples.base is ds.samples)
assert_equal(len(split.sa['chunks'].unique), 1)
assert_true('lastsplit' in split.a)
assert_true(splits[-1].a.lastsplit)
# now again, more customized
spl2 = Splitter('targets', attr_values = [0,1,1,2,3,3,3], count=4,
noslicing=True)
splits = list(spl2.generate(ds))
assert_equal(len(splits), 4)
for split in splits:
# it should NOT have perform basic slicing!
assert_false(split.samples.base is ds.samples)
assert_equal(len(split.sa['targets'].unique), 1)
assert_equal(len(split.sa['chunks'].unique), 10)
assert_true(splits[-1].a.lastsplit)
# two should be identical
assert_array_equal(splits[1].samples, splits[2].samples)
# now go wild and split by feature attribute
ds.fa['roi'] = np.repeat([0,1], 5)
# splitter should auto-detect that this is a feature attribute
spl3 = Splitter('roi')
splits = list(spl3.generate(ds))
assert_equal(len(splits), 2)
for split in splits:
assert_true(split.samples.base is ds.samples)
assert_equal(len(split.fa['roi'].unique), 1)
assert_equal(split.shape, (100, 5))
# and finally test chained splitters
cspl = ChainNode([spl2, spl3, spl1])
splits = list(cspl.generate(ds))
# 4 target splits and 2 roi splits each and 10 chunks each
assert_equal(len(splits), 80)
def test_sifter():
# somewhat duplicating the doctest
ds = Dataset(samples=np.arange(8).reshape((4,2)),
sa={'chunks': [ 0 , 1 , 2 , 3 ],
'targets': ['c', 'c', 'p', 'p']})
par = ChainNode([NFoldPartitioner(cvtype=2, attr='chunks'),
Sifter([('partitions', 2),
('targets', ['c', 'p'])])
])
dss = list(par.generate(ds))
assert_equal(len(dss), 4)
for ds_ in dss:
testing = ds[ds_.sa.partitions == 2]
assert_array_equal(np.unique(testing.sa.targets), ['c', 'p'])
# and we still have both targets present in training
training = ds[ds_.sa.partitions == 1]
assert_array_equal(np.unique(training.sa.targets), ['c', 'p'])
def test_sifter():
# somewhat duplicating the doctest
ds = Dataset(samples=np.arange(8).reshape((4, 2)), sa={"chunks": [0, 1, 2, 3], "targets": ["c", "c", "p", "p"]})
for sift_targets_definition in (["c", "p"], dict(uvalues=["c", "p"])):
par = ChainNode(
[
NFoldPartitioner(cvtype=2, attr="chunks"),
Sifter([("partitions", 2), ("targets", sift_targets_definition)]),
]
)
dss = list(par.generate(ds))
assert_equal(len(dss), 4)
for ds_ in dss:
testing = ds[ds_.sa.partitions == 2]
assert_array_equal(np.unique(testing.sa.targets), ["c", "p"])
# and we still have both targets present in training
training = ds[ds_.sa.partitions == 1]
assert_array_equal(np.unique(training.sa.targets), ["c", "p"])
def test_discarded_boundaries(self):
ds = datasets["hollow"]
# four runs
ds.sa["chunks"] = np.repeat(np.arange(4), 10)
# do odd even splitting for lots of boundaries in few splits
part = ChainNode([OddEvenPartitioner(), StripBoundariesSamples("chunks", 1, 2)])
parts = [d.samples.sid for d in part.generate(ds)]
# both dataset should have the same samples, because the boundaries are
# identical and the same sample should be stripped
assert_array_equal(parts[0], parts[1])
# we strip 3 samples per boundary
assert_equal(len(parts[0]), len(ds) - (3 * 3))
for i in [9, 10, 11, 19, 20, 21, 29, 30, 31]:
assert_false(i in parts[0])
def test_sifter_superord_usecase():
from mvpa2.misc.data_generators import normal_feature_dataset
from mvpa2.clfs.svm import LinearCSVMC # fast one to use for tests
from mvpa2.measures.base import CrossValidation
from mvpa2.base.node import ChainNode
from mvpa2.generators.partition import NFoldPartitioner
from mvpa2.generators.base import Sifter
# Let's simulate the beast -- 6 categories total groupped into 3
# super-ordinate, and actually without any 'superordinate' effect
# since subordinate categories independent
ds = normal_feature_dataset(
nlabels=6,
snr=100, # pure signal! ;)
perlabel=30,
nfeatures=6,
nonbogus_features=range(6),
nchunks=5)
ds.sa['subord'] = ds.sa.targets.copy()
ds.sa['superord'] = ['super%d' % (int(i[1]) % 3, )
for i in ds.targets] # 3 superord categories
# let's override original targets just to be sure that we aren't relying on them
ds.targets[:] = 0
npart = ChainNode(
[
## so we split based on superord
NFoldPartitioner(len(ds.sa['superord'].unique), attr='subord'),
## so it should select only those splits where we took 1 from
## each of the superord categories leaving things in balance
Sifter([('partitions', 2),
('superord', {
'uvalues': ds.sa['superord'].unique,
'balanced': True
})]),
],
space='partitions')
# and then do your normal where clf is space='superord'
clf = LinearCSVMC(space='superord')
cvte_regular = CrossValidation(clf,
NFoldPartitioner(),
errorfx=lambda p, t: np.mean(p == t))
cvte_super = CrossValidation(clf,
npart,
errorfx=lambda p, t: np.mean(p == t))
accs_regular = cvte_regular(ds)
accs_super = cvte_super(ds)
# With sifting we should get only 2^3 = 8 splits
assert (len(accs_super) == 8)
# I don't think that this would ever fail, so not marking it labile
assert (np.mean(accs_regular) > .8)
assert (np.mean(accs_super) < .6)
def test_discarded_boundaries(self):
ds = datasets['hollow']
# four runs
ds.sa['chunks'] = np.repeat(np.arange(4), 10)
# do odd even splitting for lots of boundaries in few splits
part = ChainNode([OddEvenPartitioner(),
StripBoundariesSamples('chunks', 1, 2)])
parts = [d.samples.sid for d in part.generate(ds)]
# both dataset should have the same samples, because the boundaries are
# identical and the same sample should be stripped
assert_array_equal(parts[0], parts[1])
# we strip 3 samples per boundary
assert_equal(len(parts[0]), len(ds) - (3 * 3))
for i in [9, 10, 11, 19, 20, 21, 29, 30, 31]:
assert_false(i in parts[0])
def test_sifter():
# somewhat duplicating the doctest
ds = Dataset(samples=np.arange(8).reshape((4,2)),
sa={'chunks': [ 0 , 1 , 2 , 3 ],
'targets': ['c', 'c', 'p', 'p']})
for sift_targets_definition in (['c', 'p'],
dict(uvalues=['c', 'p'])):
par = ChainNode([NFoldPartitioner(cvtype=2, attr='chunks'),
Sifter([('partitions', 2),
('targets', sift_targets_definition)])
])
dss = list(par.generate(ds))
assert_equal(len(dss), 4)
for ds_ in dss:
testing = ds[ds_.sa.partitions == 2]
assert_array_equal(np.unique(testing.sa.targets), ['c', 'p'])
# and we still have both targets present in training
training = ds[ds_.sa.partitions == 1]
assert_array_equal(np.unique(training.sa.targets), ['c', 'p'])
def test_exclude_targets_combinations():
partitioner = ChainNode(
[NFoldPartitioner(), ExcludeTargetsCombinationsPartitioner(k=2, targets_attr="targets", space="partitions")],
space="partitions",
)
from mvpa2.misc.data_generators import normal_feature_dataset
ds = normal_feature_dataset(snr=0.0, nlabels=4, perlabel=3, nchunks=3, nonbogus_features=[0, 1, 2, 3], nfeatures=4)
partitions = list(partitioner.generate(ds))
assert_equal(len(partitions), 3 * 6)
splitter = Splitter("partitions")
combs = []
comb_chunks = []
for p in partitions:
trds, teds = list(splitter.generate(p))[:2]
comb = tuple(np.unique(teds.targets))
combs.append(comb)
comb_chunks.append(comb + tuple(np.unique(teds.chunks)))
assert_equal(len(set(combs)), 6) # just 6 possible combinations of 2 out of 4
assert_equal(len(set(comb_chunks)), 3 * 6) # all unique
def test_sifter_with_balancing():
# extended previous test which was already
# "... somewhat duplicating the doctest"
ds = Dataset(samples=np.arange(12).reshape((-1, 2)),
sa={
'chunks': [0, 1, 2, 3, 4, 5],
'targets': ['c', 'c', 'c', 'p', 'p', 'p']
})
# Without sifter -- just to assure that we do get all of them
# i.e. 6*5*4*3/(4!) = 15
par = ChainNode([NFoldPartitioner(cvtype=4, attr='chunks')])
assert_equal(len(list(par.generate(ds))), 15)
# so we will take 4 chunks out of available 7, but would care only
# about those partitions where we have balanced number of 'c' and 'p'
# entries
assert_raises(
ValueError,
lambda x: list(Sifter([('targets', dict(wrong=1))]).generate(x)), ds)
par = ChainNode([
NFoldPartitioner(cvtype=4, attr='chunks'),
Sifter([('partitions', 2),
('targets', dict(uvalues=['c', 'p'], balanced=True))])
])
dss = list(par.generate(ds))
# print [ x[x.sa.partitions==2].sa.targets for x in dss ]
assert_equal(len(dss), 9)
for ds_ in dss:
testing = ds[ds_.sa.partitions == 2]
assert_array_equal(np.unique(testing.sa.targets), ['c', 'p'])
# and we still have both targets present in training
training = ds[ds_.sa.partitions == 1]
assert_array_equal(np.unique(training.sa.targets), ['c', 'p'])
def test_split_clf_on_chainpartitioner(self):
# pretty much a smoke test for #156
ds = datasets['uni2small']
part = ChainNode([NFoldPartitioner(cvtype=1),
Balancer(attr='targets', count=2,
limit='partitions', apply_selection=True)])
partitions = list(part.generate(ds))
sclf = SplitClassifier(sample_clf_lin, part, enable_ca=['stats', 'splits'])
sclf.train(ds)
pred = sclf.predict(ds)
assert_equal(len(pred), len(ds)) # rudimentary check
assert_equal(len(sclf.ca.splits), len(partitions))
assert_equal(len(sclf.clfs), len(partitions))
# now let's do sensitivity analyzer just in case
sclf.untrain()
sensana = sclf.get_sensitivity_analyzer()
sens = sensana(ds)
# basic check that sensitivities varied across splits
from mvpa2.mappers.fx import FxMapper
sens_stds = FxMapper('samples', np.std, uattrs=['targets'])(sens)
assert_true(np.any(sens_stds != 0))
def test_sifter_superord_usecase():
from mvpa2.misc.data_generators import normal_feature_dataset
from mvpa2.clfs.svm import LinearCSVMC # fast one to use for tests
from mvpa2.measures.base import CrossValidation
from mvpa2.base.node import ChainNode
from mvpa2.generators.partition import NFoldPartitioner
from mvpa2.generators.base import Sifter
ds = _get_superord_dataset()
npart = ChainNode(
[
## so we split based on superord
NFoldPartitioner(len(ds.sa['superord'].unique), attr='subord'),
## so it should select only those splits where we took 1 from
## each of the superord categories leaving things in balance
Sifter([('partitions', 2),
('superord', {
'uvalues': ds.sa['superord'].unique,
'balanced': True
})]),
],
space='partitions')
# and then do your normal where clf is space='superord'
clf = LinearCSVMC(space='superord')
cvte_regular = CrossValidation(clf,
NFoldPartitioner(),
errorfx=lambda p, t: np.mean(p == t))
cvte_super = CrossValidation(clf,
npart,
errorfx=lambda p, t: np.mean(p == t))
accs_regular = cvte_regular(ds)
accs_super = cvte_super(ds)
# With sifting we should get only 2^3 = 8 splits
assert (len(accs_super) == 8)
# I don't think that this would ever fail, so not marking it labile
assert (np.mean(accs_regular) > .8)
assert (np.mean(accs_super) < .6)
def test_compound_node(self):
data = np.asarray([[1, 2, 3, 4]], dtype=np.float_).T
ds = AttrDataset(data, sa=dict(targets=[0, 0, 1, 1]))
add = lambda x: lambda y: x + y
mul = lambda x: lambda y: x * y
add2 = FxNode(add(2))
mul3 = FxNode(mul(3))
assert_array_equal(add2(ds).samples, data + 2)
add2mul3 = ChainNode([add2, mul3])
assert_array_equal(add2mul3(ds), (data + 2) * 3)
add2_mul3v = CombinedNode([add2, mul3], 'v')
add2_mul3h = CombinedNode([add2, mul3], 'h')
assert_array_equal(
add2_mul3v(ds).samples, np.vstack((data + 2, data * 3)))
assert_array_equal(
add2_mul3h(ds).samples, np.hstack((data + 2, data * 3)))
def test_confusion_as_node():
from mvpa2.misc.data_generators import normal_feature_dataset
from mvpa2.clfs.gnb import GNB
from mvpa2.clfs.transerror import Confusion
ds = normal_feature_dataset(snr=2.0,
perlabel=42,
nchunks=3,
nonbogus_features=[0, 1],
nfeatures=2)
clf = GNB()
cv = CrossValidation(clf,
NFoldPartitioner(),
errorfx=None,
postproc=Confusion(labels=ds.UT),
enable_ca=['stats'])
res = cv(ds)
# needs to be identical to CA
assert_array_equal(res.samples, cv.ca.stats.matrix)
assert_array_equal(res.sa.predictions, ds.UT)
assert_array_equal(res.fa.targets, ds.UT)
skip_if_no_external('scipy')
from mvpa2.clfs.transerror import BayesConfusionHypothesis
from mvpa2.base.node import ChainNode
# same again, but this time with Bayesian hypothesis testing at the end
cv = CrossValidation(clf,
NFoldPartitioner(),
errorfx=None,
postproc=ChainNode((Confusion(labels=ds.UT),
BayesConfusionHypothesis())))
res = cv(ds)
# only two possible hypothesis with two classes
assert_equals(len(res), 2)
# the first hypothesis is the can't discriminate anything
assert_equal(len(res.sa.hypothesis[0]), 1)
assert_equal(len(res.sa.hypothesis[0][0]), 2)
# and the hypothesis is actually less likely than the other one
# (both classes can be distinguished)
assert (np.e**res.samples[0, 0] < np.e**res.samples[1, 0])
def get_crossvalidation_instance(learner, partitioner, errorfx,
sampling_repetitions=1,
learner_space='targets',
balance_training=None,
permutations=0,
avg_datafold_results=True,
prob_tail='left'):
from mvpa2.base.node import ChainNode
from mvpa2.measures.base import CrossValidation
if not balance_training is None:
# balance training data
try:
amount = int(balance_training)
except ValueError:
try:
amount = float(balance_training)
except ValueError:
amount = balance_training
from mvpa2.generators.resampling import Balancer
balancer = Balancer(amount=amount, attr=learner_space,
count=sampling_repetitions,
limit={partitioner.get_space(): 1},
apply_selection=True,
include_offlimit=True)
else:
balancer = None
# set learner space
learner.set_space(learner_space)
# setup generator for data folding -- put in a chain node for easy
# amending
gennode = ChainNode([partitioner], space=partitioner.get_space())
if avg_datafold_results:
from mvpa2.mappers.fx import mean_sample
postproc = mean_sample()
else:
postproc = None
if not balancer is None:
# enable balancing step for each partitioning step
gennode.append(balancer)
if permutations > 0:
from mvpa2.generators.base import Repeater
from mvpa2.generators.permutation import AttributePermutator
from mvpa2.clfs.stats import MCNullDist
# how often do we want to shuffle the data
repeater = Repeater(count=permutations)
# permute the training part of a dataset exactly ONCE
permutator = AttributePermutator(
learner_space,
limit={partitioner.get_space(): 1},
count=1)
# CV with null-distribution estimation that permutes the training data for
# each fold independently
perm_gen_node = copy.deepcopy(gennode)
perm_gen_node.append(permutator)
null_cv = CrossValidation(learner,
perm_gen_node,
postproc=postproc,
errorfx=errorfx)
# Monte Carlo distribution estimator
distr_est = MCNullDist(repeater,
tail=prob_tail,
measure=null_cv,
enable_ca=['dist_samples'])
# pass the p-values as feature attributes on to the results
pass_attr = [('ca.null_prob', 'fa', 1)]
else:
distr_est = None
pass_attr = None
# final CV node
cv = CrossValidation(learner,
gennode,
errorfx=errorfx,
null_dist=distr_est,
postproc=postproc,
enable_ca=['stats', 'null_prob'],
pass_attr=pass_attr)
return cv
def group_sample_loser_measure(attrs=('targets', )):
'''takes loser after meaning over attrs'''
return ChainNode((mean_group_sample(attrs), sample_loser_measure()))
def plot_feature_hist(dataset,
xlim=None,
noticks=True,
targets_attr='targets',
chunks_attr=None,
**kwargs):
"""Plot histograms of feature values for each labels.
Parameters
----------
dataset : Dataset
xlim : None or 2-tuple
Common x-axis limits for all histograms.
noticks : bool
If True, no axis ticks will be plotted. This is useful to save
space in large plots.
targets_attr : string, optional
Name of samples attribute to be used as targets
chunks_attr : None or string
If a string, a histogram will be plotted per each target and each
chunk (as defined in sa named `chunks_attr`), resulting is a
histogram grid (targets x chunks).
**kwargs
Any additional arguments are passed to matplotlib's hist().
"""
lsplit = ChainNode([
NFoldPartitioner(1, attr=targets_attr),
Splitter('partitions', attr_values=[2])
])
csplit = ChainNode([
NFoldPartitioner(1, attr=chunks_attr),
Splitter('partitions', attr_values=[2])
])
nrows = len(dataset.sa[targets_attr].unique)
ncols = len(dataset.sa[chunks_attr].unique)
def doplot(data):
"""Just a little helper which plots the histogram and removes
ticks etc"""
pl.hist(data, **kwargs)
if xlim is not None:
pl.xlim(xlim)
if noticks:
pl.yticks([])
pl.xticks([])
fig = 1
# for all labels
for row, ds in enumerate(lsplit.generate(dataset)):
if chunks_attr:
for col, d in enumerate(csplit.generate(ds)):
pl.subplot(nrows, ncols, fig)
doplot(d.samples.ravel())
if row == 0:
pl.title('C:' + str(d.sa[chunks_attr].unique[0]))
if col == 0:
pl.ylabel('L:' + str(d.sa[targets_attr].unique[0]))
fig += 1
else:
pl.subplot(1, nrows, fig)
doplot(ds.samples)
pl.title('L:' + str(ds.sa[targets_attr].unique[0]))
fig += 1
def test_gnbsearchlight_permutations():
import mvpa2
from mvpa2.base.node import ChainNode
from mvpa2.clfs.gnb import GNB
from mvpa2.generators.base import Repeater
from mvpa2.generators.partition import NFoldPartitioner, OddEvenPartitioner
#import mvpa2.generators.permutation
#reload(mvpa2.generators.permutation)
from mvpa2.generators.permutation import AttributePermutator
from mvpa2.testing.datasets import datasets
from mvpa2.measures.base import CrossValidation
from mvpa2.measures.gnbsearchlight import sphere_gnbsearchlight
from mvpa2.measures.searchlight import sphere_searchlight
from mvpa2.mappers.fx import mean_sample
from mvpa2.misc.errorfx import mean_mismatch_error
from mvpa2.clfs.stats import MCNullDist
from mvpa2.testing.tools import assert_raises, ok_, assert_array_less
# mvpa2.debug.active = ['APERM', 'SLC'] #, 'REPM']
# mvpa2.debug.metrics += ['pid']
count = 10
nproc = 1 + int(mvpa2.externals.exists('pprocess'))
ds = datasets['3dsmall'].copy()
ds.fa['voxel_indices'] = ds.fa.myspace
slkwargs = dict(radius=3, space='voxel_indices', enable_ca=['roi_sizes'],
center_ids=[1, 10, 70, 100])
mvpa2.seed(mvpa2._random_seed)
clf = GNB()
splt = NFoldPartitioner(cvtype=2, attr='chunks')
repeater = Repeater(count=count)
permutator = AttributePermutator('targets', limit={'partitions': 1}, count=1)
null_sl = sphere_gnbsearchlight(clf, ChainNode([splt, permutator], space=splt.get_space()),
postproc=mean_sample(), errorfx=mean_mismatch_error,
**slkwargs)
distr_est = MCNullDist(repeater, tail='left', measure=null_sl,
enable_ca=['dist_samples'])
sl = sphere_gnbsearchlight(clf, splt,
reuse_neighbors=True,
null_dist=distr_est, postproc=mean_sample(),
errorfx=mean_mismatch_error,
**slkwargs)
if __debug__: # assert is done only without -O mode
assert_raises(NotImplementedError, sl, ds)
# "ad-hoc searchlights can't handle yet varying targets across partitions"
if False:
# after above limitation is removed -- enable
sl_map = sl(ds)
sl_null_prob = sl.ca.null_prob.samples.copy()
mvpa2.seed(mvpa2._random_seed)
### 'normal' Searchlight
clf = GNB()
splt = NFoldPartitioner(cvtype=2, attr='chunks')
repeater = Repeater(count=count)
permutator = AttributePermutator('targets', limit={'partitions': 1}, count=1)
# rng=np.random.RandomState(0)) # to trigger failure since the same np.random state
# would be reused across all pprocesses
null_cv = CrossValidation(clf, ChainNode([splt, permutator], space=splt.get_space()),
postproc=mean_sample())
null_sl_normal = sphere_searchlight(null_cv, nproc=nproc, **slkwargs)
distr_est_normal = MCNullDist(repeater, tail='left', measure=null_sl_normal,
enable_ca=['dist_samples'])
cv = CrossValidation(clf, splt, errorfx=mean_mismatch_error,
enable_ca=['stats'], postproc=mean_sample() )
sl = sphere_searchlight(cv, nproc=nproc, null_dist=distr_est_normal, **slkwargs)
sl_map_normal = sl(ds)
sl_null_prob_normal = sl.ca.null_prob.samples.copy()
# For every feature -- we should get some variance in estimates In
# case of failure they are all really close to each other (up to
# numerical precision), so variance will be close to 0
assert_array_less(-np.var(distr_est_normal.ca.dist_samples.samples[0],
axis=1), -1e-5)
for s in distr_est_normal.ca.dist_samples.samples[0]:
ok_(len(np.unique(s)) > 1)
def test_factorialpartitioner():
# Test against sifter and chainmap implemented in test_usecases
# -- code below copied from test_usecases --
# Let's simulate the beast -- 6 categories total groupped into 3
# super-ordinate, and actually without any 'superordinate' effect
# since subordinate categories independent
ds = normal_feature_dataset(
nlabels=6, snr=100, perlabel=30, nfeatures=6, nonbogus_features=range(6), nchunks=5 # pure signal! ;)
)
ds.sa["subord"] = ds.sa.targets.copy()
ds.sa["superord"] = ["super%d" % (int(i[1]) % 3,) for i in ds.targets] # 3 superord categories
# let's override original targets just to be sure that we aren't relying on them
ds.targets[:] = 0
# let's make two other datasets to test later
# one superordinate category only
ds_1super = ds.copy()
ds_1super.sa["superord"] = ["super1" for i in ds_1super.targets]
# one superordinate category has only one subordinate
# ds_unbalanced = ds.copy()
# nsuper1 = np.sum(ds_unbalanced.sa.superord == 'super1')
# mask_superord = ds_unbalanced.sa.superord == 'super1'
# uniq_subord = np.unique(ds_unbalanced.sa.subord[mask_superord])
# ds_unbalanced.sa.subord[mask_superord] = [uniq_subord[0] for i in range(nsuper1)]
ds_unbalanced = Dataset(range(4), sa={"subord": [0, 0, 1, 2], "superord": [1, 1, 2, 2]})
npart = ChainNode(
[
## so we split based on superord
NFoldPartitioner(len(ds.sa["superord"].unique), attr="subord"),
## so it should select only those splits where we took 1 from
## each of the superord categories leaving things in balance
Sifter([("partitions", 2), ("superord", {"uvalues": ds.sa["superord"].unique, "balanced": True})]),
],
space="partitions",
)
# now the new implementation
factpart = FactorialPartitioner(NFoldPartitioner(attr="subord"), attr="superord")
partitions_npart = [p.sa.partitions for p in npart.generate(ds)]
partitions_factpart = [p.sa.partitions for p in factpart.generate(ds)]
assert_array_equal(np.sort(partitions_npart), np.sort(partitions_factpart))
# now let's check it behaves correctly if we have only one superord class
nfold = NFoldPartitioner(attr="subord")
partitions_nfold = [p.sa.partitions for p in nfold.generate(ds_1super)]
partitions_factpart = [p.sa.partitions for p in factpart.generate(ds_1super)]
assert_array_equal(np.sort(partitions_nfold), np.sort(partitions_factpart))
# smoke test for unbalanced subord classes
warning_msg = (
"One or more superordinate attributes do not have the same "
"number of subordinate attributes. This could yield to "
"unbalanced partitions."
)
with assert_warnings([(RuntimeWarning, warning_msg)]):
partitions_factpart = [p.sa.partitions for p in factpart.generate(ds_unbalanced)]
partitions_unbalanced = [np.array([2, 2, 2, 1]), np.array([2, 2, 1, 2])]
superord_unbalanced = [([2], [1, 1, 2]), ([2], [1, 1, 2])]
subord_unbalanced = [([2], [0, 0, 1]), ([1], [0, 0, 2])]
for out_part, true_part, super_out, sub_out in zip(
partitions_factpart, partitions_unbalanced, superord_unbalanced, subord_unbalanced
):
assert_array_equal(out_part, true_part)
assert_array_equal(
(ds_unbalanced[out_part == 1].sa.superord.tolist(), ds_unbalanced[out_part == 2].sa.superord.tolist()),
super_out,
)
assert_array_equal(
(ds_unbalanced[out_part == 1].sa.subord.tolist(), ds_unbalanced[out_part == 2].sa.subord.tolist()), sub_out
)
# now let's test on a dummy dataset
ds_dummy = Dataset(range(4), sa={"subord": range(4), "superord": [1, 2] * 2})
partitions_factpart = [p.sa.partitions for p in factpart.generate(ds_dummy)]
assert_array_equal(partitions_factpart, [[2, 2, 1, 1], [2, 1, 1, 2], [1, 2, 2, 1], [1, 1, 2, 2]])
def test_factorialpartitioner():
# Test against sifter and chainmap implemented in test_usecases
# -- code below copied from test_usecases --
# Let's simulate the beast -- 6 categories total groupped into 3
# super-ordinate, and actually without any 'superordinate' effect
# since subordinate categories independent
ds = normal_feature_dataset(
nlabels=6,
snr=100, # pure signal! ;)
perlabel=30,
nfeatures=6,
nonbogus_features=range(6),
nchunks=5)
ds.sa['subord'] = ds.sa.targets.copy()
ds.sa['superord'] = ['super%d' % (int(i[1]) % 3, )
for i in ds.targets] # 3 superord categories
# let's override original targets just to be sure that we aren't relying on them
ds.targets[:] = 0
# let's make two other datasets to test later
# one superordinate category only
ds_1super = ds.copy()
ds_1super.sa['superord'] = ['super1' for i in ds_1super.targets]
# one superordinate category has only one subordinate
#ds_unbalanced = ds.copy()
#nsuper1 = np.sum(ds_unbalanced.sa.superord == 'super1')
#mask_superord = ds_unbalanced.sa.superord == 'super1'
#uniq_subord = np.unique(ds_unbalanced.sa.subord[mask_superord])
#ds_unbalanced.sa.subord[mask_superord] = [uniq_subord[0] for i in range(nsuper1)]
ds_unbalanced = Dataset(range(4),
sa={
'subord': [0, 0, 1, 2],
'superord': [1, 1, 2, 2]
})
npart = ChainNode(
[
## so we split based on superord
NFoldPartitioner(len(ds.sa['superord'].unique), attr='subord'),
## so it should select only those splits where we took 1 from
## each of the superord categories leaving things in balance
Sifter([('partitions', 2),
('superord', {
'uvalues': ds.sa['superord'].unique,
'balanced': True
})]),
],
space='partitions')
# now the new implementation
factpart = FactorialPartitioner(NFoldPartitioner(attr='subord'),
attr='superord')
partitions_npart = [p.sa.partitions for p in npart.generate(ds)]
partitions_factpart = [p.sa.partitions for p in factpart.generate(ds)]
assert_array_equal(np.sort(partitions_npart), np.sort(partitions_factpart))
# now let's check it behaves correctly if we have only one superord class
nfold = NFoldPartitioner(attr='subord')
partitions_nfold = [p.sa.partitions for p in nfold.generate(ds_1super)]
partitions_factpart = [
p.sa.partitions for p in factpart.generate(ds_1super)
]
assert_array_equal(np.sort(partitions_nfold), np.sort(partitions_factpart))
# smoke test for unbalanced subord classes
warning_msg = 'One or more superordinate attributes do not have the same '\
'number of subordinate attributes. This could yield to '\
'unbalanced partitions.'
with assert_warnings([(RuntimeWarning, warning_msg)]):
partitions_factpart = [
p.sa.partitions for p in factpart.generate(ds_unbalanced)
]
partitions_unbalanced = [np.array([2, 2, 2, 1]), np.array([2, 2, 1, 2])]
superord_unbalanced = [([2], [1, 1, 2]), ([2], [1, 1, 2])]
subord_unbalanced = [([2], [0, 0, 1]), ([1], [0, 0, 2])]
for out_part, true_part, super_out, sub_out in \
zip(partitions_factpart, partitions_unbalanced,
superord_unbalanced, subord_unbalanced):
assert_array_equal(out_part, true_part)
assert_array_equal((ds_unbalanced[out_part == 1].sa.superord.tolist(),
ds_unbalanced[out_part == 2].sa.superord.tolist()),
super_out)
assert_array_equal((ds_unbalanced[out_part == 1].sa.subord.tolist(),
ds_unbalanced[out_part == 2].sa.subord.tolist()),
sub_out)
# now let's test on a dummy dataset
ds_dummy = Dataset(range(4),
sa={
'subord': range(4),
'superord': [1, 2] * 2
})
partitions_factpart = [
p.sa.partitions for p in factpart.generate(ds_dummy)
]
assert_array_equal(
partitions_factpart,
[[2, 2, 1, 1], [2, 1, 1, 2], [1, 2, 2, 1], [1, 1, 2, 2]])
def hist(dataset,
xgroup_attr=None,
ygroup_attr=None,
xlim=None,
ylim=None,
noticks=False,
**kwargs):
"""Compute and draw feature histograms (for groups of samples)
This is a convenience wrapper around matplotlib's hist() function. It
supports it entire API, but data is taken from an input dataset. In
addition, feature histograms for groups of dataset samples can be drawn as
an array of subplots. Using ``xgroup_attr`` and ``ygroup_attr`` up to two
sample attributes can be selected and samples groups are defined by their
unique values. For example, plotting histograms for all combinations of
``targets`` and ``chunks`` attribute values in a dataset is done by this
code:
>>> from mvpa2.viz import hist
>>> from mvpa2.misc.data_generators import normal_feature_dataset
>>> ds = normal_feature_dataset(10, 3, 10, 5)
>>> plots = hist(ds, ygroup_attr='targets', xgroup_attr='chunks',
... noticks=None, xlim=(-.5,.5), normed=True)
>>> len(plots)
15
This function can also be used with plain arrays, in which case it will
fall back on the behavior of matplotlib's hist() and additional
functionality is not available.
Parameters
----------
dataset : Dataset or array
xgroup_attr : string, optional
Name of a samples attribute to be used as targets
ygroup_attr : None or string, optional
If a string, a histogram will be plotted per each target and each
chunk (as defined in sa named `chunks_attr`), resulting is a
histogram grid (targets x chunks).
xlim : None or 2-tuple, optional
Common x-axis limits for all histograms.
ylim : None or 2-tuple, optional
Common y-axis limits for all histograms.
noticks : bool or None, optional
If True, no axis ticks will be plotted. If False, each histogram subplot
will have its own ticks. If None, only the outer subplots will
have ticks. This is useful to save space in large plots, but should be
combined with ``xlim`` and ``ylim`` arguments in order to ensure equal
axes across subplots.
**kwargs
Any additional arguments are passed to matplotlib's hist().
Returns
-------
list
List of figure handlers for all generated subplots.
"""
xgroup = {'attr': xgroup_attr}
ygroup = {'attr': ygroup_attr}
for grp in (xgroup, ygroup):
if grp['attr'] is not None and is_datasetlike(dataset):
grp['split'] = ChainNode([
NFoldPartitioner(1, attr=grp['attr']),
Splitter('partitions', attr_values=[2])
])
grp['gen'] = lambda s, x: s.generate(x)
grp['npanels'] = len(dataset.sa[grp['attr']].unique)
else:
grp['split'] = None
grp['gen'] = lambda s, x: [x]
grp['npanels'] = 1
fig = 1
nrows = ygroup['npanels']
ncols = xgroup['npanels']
subplots = []
# for all labels
for row, ds in enumerate(ygroup['gen'](ygroup['split'], dataset)):
for col, d in enumerate(xgroup['gen'](xgroup['split'], ds)):
ax = pl.subplot(nrows, ncols, fig)
if is_datasetlike(d):
data = d.samples
else:
data = d
ax.hist(data.ravel(), **kwargs)
if xlim is not None:
pl.xlim(xlim)
if noticks is True or (noticks is None and row < nrows - 1):
pl.xticks([])
if noticks is True or (noticks is None and col > 0):
pl.yticks([])
if ncols > 1 and row == 0:
pl.title(str(d.sa[xgroup['attr']].unique[0]))
if nrows > 1 and col == 0:
pl.ylabel(str(d.sa[ygroup['attr']].unique[0]))
fig += 1
subplots.append(ax)
return subplots
def _call(self, ds):
return ChainNode._call(self, ds)
def get_crossvalidation_instance(learner,
partitioner,
errorfx,
sampling_repetitions=1,
learner_space='targets',
balance_training=None,
permutations=0,
avg_datafold_results=True,
prob_tail='left'):
from mvpa2.base.node import ChainNode
from mvpa2.measures.base import CrossValidation
if not balance_training is None:
# balance training data
try:
amount = int(balance_training)
except ValueError:
try:
amount = float(balance_training)
except ValueError:
amount = balance_training
from mvpa2.generators.resampling import Balancer
balancer = Balancer(amount=amount,
attr=learner_space,
count=sampling_repetitions,
limit={partitioner.get_space(): 1},
apply_selection=True,
include_offlimit=True)
else:
balancer = None
# set learner space
learner.set_space(learner_space)
# setup generator for data folding -- put in a chain node for easy
# amending
gennode = ChainNode([partitioner], space=partitioner.get_space())
if avg_datafold_results:
from mvpa2.mappers.fx import mean_sample
postproc = mean_sample()
else:
postproc = None
if not balancer is None:
# enable balancing step for each partitioning step
gennode.append(balancer)
if permutations > 0:
from mvpa2.generators.base import Repeater
from mvpa2.generators.permutation import AttributePermutator
from mvpa2.clfs.stats import MCNullDist
# how often do we want to shuffle the data
repeater = Repeater(count=permutations)
# permute the training part of a dataset exactly ONCE
permutator = AttributePermutator(learner_space,
limit={partitioner.get_space(): 1},
count=1)
# CV with null-distribution estimation that permutes the training data for
# each fold independently
perm_gen_node = copy.deepcopy(gennode)
perm_gen_node.append(permutator)
null_cv = CrossValidation(learner,
perm_gen_node,
postproc=postproc,
errorfx=errorfx)
# Monte Carlo distribution estimator
distr_est = MCNullDist(repeater,
tail=prob_tail,
measure=null_cv,
enable_ca=['dist_samples'])
# pass the p-values as feature attributes on to the results
pass_attr = [('ca.null_prob', 'fa', 1)]
else:
distr_est = None
pass_attr = None
# final CV node
cv = CrossValidation(learner,
gennode,
errorfx=errorfx,
null_dist=distr_est,
postproc=postproc,
enable_ca=['stats', 'null_prob'],
pass_attr=pass_attr)
return cv
def test_confusion_as_node():
from mvpa2.misc.data_generators import normal_feature_dataset
from mvpa2.clfs.gnb import GNB
from mvpa2.clfs.transerror import Confusion
ds = normal_feature_dataset(snr=2.0, perlabel=42, nchunks=3,
nonbogus_features=[0,1], nfeatures=2)
clf = GNB()
cv = CrossValidation(
clf, NFoldPartitioner(),
errorfx=None,
postproc=Confusion(labels=ds.UT),
enable_ca=['stats'])
res = cv(ds)
# needs to be identical to CA
assert_array_equal(res.samples, cv.ca.stats.matrix)
assert_array_equal(res.sa.predictions, ds.UT)
assert_array_equal(res.fa.targets, ds.UT)
skip_if_no_external('scipy')
from mvpa2.clfs.transerror import BayesConfusionHypothesis
from mvpa2.base.node import ChainNode
# same again, but this time with Bayesian hypothesis testing at the end
cv = CrossValidation(
clf, NFoldPartitioner(),
errorfx=None,
postproc=ChainNode([Confusion(labels=ds.UT),
BayesConfusionHypothesis()]))
res = cv(ds)
# only two possible hypothesis with two classes
assert_equals(len(res), 2)
# the first hypothesis is the can't discriminate anything
assert_equal(len(res.sa.hypothesis[0]), 1)
assert_equal(len(res.sa.hypothesis[0][0]), 2)
# and the hypothesis is actually less likely than the other one
# (both classes can be distinguished)
assert(np.e**res.samples[0,0] < np.e**res.samples[1,0])
# Let's see how well it would work within the searchlight when we also
# would like to store the hypotheses per each voxel
# Somewhat an ad-hoc solution for the answer posted on the ML
#
# run 1d searchlight of radii 0, for that just provide a .fa with coordinates
ds.fa['voxel_indices'] = [[0], [1]]
# and a custom Node which would collect .sa.hypothesis to place together along
# with the posterior probabilities
from mvpa2.base.node import Node
from mvpa2.measures.searchlight import sphere_searchlight
class KeepBothPosteriorAndHypothesis(Node):
def _call(self, ds):
out = np.zeros(1, dtype=object)
out[0] = (ds.samples, ds.sa.hypothesis)
return out
cv.postproc.append(KeepBothPosteriorAndHypothesis())
sl = sphere_searchlight(cv, radius=0, nproc=1)
res = sl(ds)
assert_equal(res.shape, (1, 2))
assert_equal(len(res.samples[0,0]), 2)
assert_equal(res.samples[0,0][0].shape, (2, 2)) # posteriors per 1st SL
assert_equal(len(res.samples[0,0][1]), 2) # 2 of hypotheses
def plot_feature_hist(dataset, xlim=None, noticks=True,
targets_attr='targets', chunks_attr=None,
**kwargs):
"""Plot histograms of feature values for each labels.
Parameters
----------
dataset : Dataset
xlim : None or 2-tuple
Common x-axis limits for all histograms.
noticks : bool
If True, no axis ticks will be plotted. This is useful to save
space in large plots.
targets_attr : string, optional
Name of samples attribute to be used as targets
chunks_attr : None or string
If a string, a histogram will be plotted per each target and each
chunk (as defined in sa named `chunks_attr`), resulting is a
histogram grid (targets x chunks).
**kwargs
Any additional arguments are passed to matplotlib's hist().
"""
lsplit = ChainNode([NFoldPartitioner(1, attr=targets_attr),
Splitter('partitions', attr_values=[2])])
csplit = ChainNode([NFoldPartitioner(1, attr=chunks_attr),
Splitter('partitions', attr_values=[2])])
nrows = len(dataset.sa[targets_attr].unique)
ncols = len(dataset.sa[chunks_attr].unique)
def doplot(data):
"""Just a little helper which plots the histogram and removes
ticks etc"""
pl.hist(data, **kwargs)
if xlim is not None:
pl.xlim(xlim)
if noticks:
pl.yticks([])
pl.xticks([])
fig = 1
# for all labels
for row, ds in enumerate(lsplit.generate(dataset)):
if chunks_attr:
for col, d in enumerate(csplit.generate(ds)):
pl.subplot(nrows, ncols, fig)
doplot(d.samples.ravel())
if row == 0:
pl.title('C:' + str(d.sa[chunks_attr].unique[0]))
if col == 0:
pl.ylabel('L:' + str(d.sa[targets_attr].unique[0]))
fig += 1
else:
pl.subplot(1, nrows, fig)
doplot(ds.samples)
pl.title('L:' + str(ds.sa[targets_attr].unique[0]))
fig += 1
def _call(self, ds):
return ChainNode._call(self, ds)
def test_factorialpartitioner():
# Test against sifter and chainmap implemented in test_usecases
# -- code below copied from test_usecases --
# Let's simulate the beast -- 6 categories total groupped into 3
# super-ordinate, and actually without any 'superordinate' effect
# since subordinate categories independent
ds = normal_feature_dataset(
nlabels=6,
snr=100, # pure signal! ;)
perlabel=30,
nfeatures=6,
nonbogus_features=range(6),
nchunks=5)
ds.sa['subord'] = ds.sa.targets.copy()
ds.sa['superord'] = ['super%d' % (int(i[1]) % 3, )
for i in ds.targets] # 3 superord categories
# let's override original targets just to be sure that we aren't relying on them
ds.targets[:] = 0
# let's make two other datasets to test later
# one superordinate category only
ds_1super = ds.copy()
ds_1super.sa['superord'] = ['super1' for i in ds_1super.targets]
# one superordinate category has only one subordinate
#ds_unbalanced = ds.copy()
#nsuper1 = np.sum(ds_unbalanced.sa.superord == 'super1')
#mask_superord = ds_unbalanced.sa.superord == 'super1'
#uniq_subord = np.unique(ds_unbalanced.sa.subord[mask_superord])
#ds_unbalanced.sa.subord[mask_superord] = [uniq_subord[0] for i in range(nsuper1)]
ds_unbalanced = Dataset(range(4),
sa={
'subord': [0, 0, 1, 2],
'superord': [1, 1, 2, 2]
})
npart = ChainNode(
[
## so we split based on superord
NFoldPartitioner(len(ds.sa['superord'].unique), attr='subord'),
## so it should select only those splits where we took 1 from
## each of the superord categories leaving things in balance
Sifter([('partitions', 2),
('superord', {
'uvalues': ds.sa['superord'].unique,
'balanced': True
})]),
],
space='partitions')
def partition(partitioner, ds_=ds):
return [p.sa.partitions for p in partitioner.generate(ds_)]
# now the new implementation
# common kwargs
factkw = dict(partitioner=NFoldPartitioner(attr='subord'), attr='superord')
fpart = FactorialPartitioner(**factkw)
p_npart = partition(npart)
p_fpart = partition(fpart)
assert_array_equal(np.sort(p_npart), np.sort(p_fpart))
fpart2 = FactorialPartitioner(count=2,
selection_strategy='first',
**factkw)
p_fpart2 = partition(fpart2)
assert_equal(len(p_fpart), 8)
assert_equal(len(p_fpart2), 2)
assert_array_equal(p_fpart[:2], p_fpart2)
# 1 equidistant -- should be the first one
fpart1 = FactorialPartitioner(count=1, **factkw)
p_fpart1 = partition(fpart1)
assert_equal(len(p_fpart1), 1)
assert_array_equal(p_fpart[:1], p_fpart1)
# 2 equidistant
fpart2 = FactorialPartitioner(count=2, **factkw)
p_fpart2 = partition(fpart2)
assert_equal(len(p_fpart2), 2)
assert_array_equal(p_fpart[::4], p_fpart2)
# without count -- should be all of them in original order
fpartr = FactorialPartitioner(selection_strategy='random', **factkw)
assert_array_equal(p_fpart, partition(fpartr))
# but if with a count we should get some selection
fpartr2 = FactorialPartitioner(selection_strategy='random',
count=2,
**factkw)
# Let's generate a number of random selections:
rand2_partitions = [partition(fpartr2) for i in xrange(10)]
for p in rand2_partitions:
assert_equal(len(p), 2)
# majority of them must be different
assert len(set([tuple(map(tuple, x)) for x in rand2_partitions])) >= 5
# now let's check it behaves correctly if we have only one superord class
nfold = NFoldPartitioner(attr='subord')
p_nfold = partition(nfold, ds_1super)
p_fpart = partition(fpart, ds_1super)
assert_array_equal(np.sort(p_nfold), np.sort(p_fpart))
# smoke test for unbalanced subord classes
warning_msg = 'One or more superordinate attributes do not have the same '\
'number of subordinate attributes. This could yield to '\
'unbalanced partitions.'
with assert_warnings([(RuntimeWarning, warning_msg)]):
p_fpart = partition(fpart, ds_unbalanced)
p_unbalanced = [np.array([2, 2, 2, 1]), np.array([2, 2, 1, 2])]
superord_unbalanced = [([2], [1, 1, 2]), ([2], [1, 1, 2])]
subord_unbalanced = [([2], [0, 0, 1]), ([1], [0, 0, 2])]
for out_part, true_part, super_out, sub_out in \
zip(p_fpart, p_unbalanced,
superord_unbalanced, subord_unbalanced):
assert_array_equal(out_part, true_part)
assert_array_equal((ds_unbalanced[out_part == 1].sa.superord.tolist(),
ds_unbalanced[out_part == 2].sa.superord.tolist()),
super_out)
assert_array_equal((ds_unbalanced[out_part == 1].sa.subord.tolist(),
ds_unbalanced[out_part == 2].sa.subord.tolist()),
sub_out)
# now let's test on a dummy dataset
ds_dummy = Dataset(range(4),
sa={
'subord': range(4),
'superord': [1, 2] * 2
})
p_fpart = partition(fpart, ds_dummy)
assert_array_equal(
p_fpart, [[2, 2, 1, 1], [2, 1, 1, 2], [1, 2, 2, 1], [1, 1, 2, 2]])