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_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_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_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_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_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 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]])