def test_custom_split(self):
#simulate half splitter
hs = CustomPartitioner([(None,[0,1,2,3,4]),(None,[5,6,7,8,9])])
spl = Splitter(attr='partitions')
splits = [ list(spl.generate(p)) for p in hs.generate(self.data) ]
self.failUnless(len(splits) == 2)
for i,p in enumerate(splits):
self.failUnless( len(p) == 2 )
self.failUnless( p[0].nsamples == 50 )
self.failUnless( p[1].nsamples == 50 )
assert_array_equal(splits[0][1].sa['chunks'].unique, [0, 1, 2, 3, 4])
assert_array_equal(splits[0][0].sa['chunks'].unique, [5, 6, 7, 8, 9])
assert_array_equal(splits[1][1].sa['chunks'].unique, [5, 6, 7, 8, 9])
assert_array_equal(splits[1][0].sa['chunks'].unique, [0, 1, 2, 3, 4])
# check fully customized split with working and validation set specified
cs = CustomPartitioner([([0,3,4],[5,9])])
# we want to discared the unselected partition of the data, hence attr_value
spl = Splitter(attr='partitions', attr_values=[1,2])
splits = [ list(spl.generate(p)) for p in cs.generate(self.data) ]
self.failUnless(len(splits) == 1)
for i,p in enumerate(splits):
self.failUnless( len(p) == 2 )
self.failUnless( p[0].nsamples == 30 )
self.failUnless( p[1].nsamples == 20 )
self.failUnless((splits[0][1].sa['chunks'].unique == [5, 9]).all())
self.failUnless((splits[0][0].sa['chunks'].unique == [0, 3, 4]).all())
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_slicing(self):
hs = HalfPartitioner()
spl = Splitter(attr='partitions')
splits = list(hs.generate(self.data))
for s in splits:
# partitioned dataset shared the data
assert_true(s.samples.base is self.data.samples)
splits = [ list(spl.generate(p)) for p in hs.generate(self.data) ]
for s in splits:
# we get slicing all the time
assert_true(s[0].samples.base.base is self.data.samples)
assert_true(s[1].samples.base.base is self.data.samples)
spl = Splitter(attr='partitions', noslicing=True)
splits = [ list(spl.generate(p)) for p in hs.generate(self.data) ]
for s in splits:
# we no slicing at all
assert_false(s[0].samples.base is self.data.samples)
assert_false(s[1].samples.base is self.data.samples)
nfs = NFoldPartitioner()
spl = Splitter(attr='partitions')
splits = [ list(spl.generate(p)) for p in nfs.generate(self.data) ]
for i, s in enumerate(splits):
# training only first and last split
if i == 0 or i == len(splits) - 1:
assert_true(s[0].samples.base.base is self.data.samples)
else:
assert_true(s[0].samples.base is None)
# we get slicing all the time
assert_true(s[1].samples.base.base is self.data.samples)
step_ds = Dataset(np.random.randn(20,2),
sa={'chunks': np.tile([0,1], 10)})
oes = OddEvenPartitioner()
spl = Splitter(attr='partitions')
splits = list(oes.generate(step_ds))
for s in splits:
# partitioned dataset shared the data
assert_true(s.samples.base is step_ds.samples)
splits = [ list(spl.generate(p)) for p in oes.generate(step_ds) ]
assert_equal(len(splits), 2)
for s in splits:
# we get slicing all the time
assert_true(s[0].samples.base.base is step_ds.samples)
assert_true(s[1].samples.base.base is step_ds.samples)
def test_label_splitter(self):
oes = OddEvenPartitioner(attr='targets')
spl = Splitter(attr='partitions')
splits = [ list(spl.generate(p)) for p in oes.generate(self.data) ]
assert_array_equal(splits[0][0].sa['targets'].unique, [0,2])
assert_array_equal(splits[0][1].sa['targets'].unique, [1,3])
assert_array_equal(splits[1][0].sa['targets'].unique, [1,3])
assert_array_equal(splits[1][1].sa['targets'].unique, [0,2])
def test_simplest_cv_pat_gen(self):
# create the generator
nfs = NFoldPartitioner(cvtype=1)
spl = Splitter(attr='partitions')
# now get the xval pattern sets One-Fold CV)
xvpat = [ list(spl.generate(p)) for p in nfs.generate(self.data) ]
self.failUnless( len(xvpat) == 10 )
for i,p in enumerate(xvpat):
self.failUnless( len(p) == 2 )
self.failUnless( p[0].nsamples == 90 )
self.failUnless( p[1].nsamples == 10 )
self.failUnless( p[1].chunks[0] == i )
def test_half_split(self):
hs = HalfPartitioner()
spl = Splitter(attr='partitions')
splits = [ list(spl.generate(p)) for p in hs.generate(self.data) ]
self.failUnless(len(splits) == 2)
for i,p in enumerate(splits):
self.failUnless( len(p) == 2 )
self.failUnless( p[0].nsamples == 50 )
self.failUnless( p[1].nsamples == 50 )
assert_array_equal(splits[0][1].sa['chunks'].unique, [0, 1, 2, 3, 4])
assert_array_equal(splits[0][0].sa['chunks'].unique, [5, 6, 7, 8, 9])
assert_array_equal(splits[1][1].sa['chunks'].unique, [5, 6, 7, 8, 9])
assert_array_equal(splits[1][0].sa['chunks'].unique, [0, 1, 2, 3, 4])
# check if it works on pure odd and even chunk ids
moresplits = [ list(spl.generate(p)) for p in hs.generate(splits[0][0])]
for split in moresplits:
self.failUnless(split[0] != None)
self.failUnless(split[1] != None)
def test_repeated_features(self):
print self.dataset
print self.dataset.fa.nonbogus_targets
class CountFeatures(Measure):
is_trained = True
def _call(self, ds):
return ds.nfeatures
cf = CountFeatures()
spl = Splitter('fa.nonbogus_targets')
nsplits = len(list(spl.generate(self.dataset)))
assert_equal(nsplits, 3)
rm = RepeatedMeasure(cf, spl, concat_as='features')
res = rm(self.dataset)
assert_equal(res.shape, (1, nsplits))
assert_array_equal(res.samples[0], [18,1,1])
def test_counted_splitting(self):
spl = Splitter(attr='partitions')
# count > #chunks, should result in 10 splits
nchunks = len(self.data.sa['chunks'].unique)
for strategy in Partitioner._STRATEGIES:
for count, target in [ (nchunks*2, nchunks),
(nchunks, nchunks),
(nchunks-1, nchunks-1),
(3, 3),
(0, 0),
(1, 1)
]:
nfs = NFoldPartitioner(cvtype=1, count=count,
selection_strategy=strategy)
splits = [ list(spl.generate(p)) for p in nfs.generate(self.data) ]
self.failUnless(len(splits) == target)
chosenchunks = [int(s[1].uniquechunks) for s in splits]
# Test if configuration matches as well
nsplits_cfg = len(nfs.get_partition_specs(self.data))
self.failUnlessEqual(nsplits_cfg, target)
# Check if "lastsplit" dsattr was assigned appropriately
nsplits = len(splits)
if nsplits > 0:
# dummy-proof testing of last split
for ds_ in splits[-1]:
self.failUnless(ds_.a.lastpartitionset)
# test all now
for isplit,split in enumerate(splits):
for ds_ in split:
ds_.a.lastpartitionset == isplit==nsplits-1
# Check results of different strategies
if strategy == 'first':
self.failUnlessEqual(chosenchunks, range(target))
elif strategy == 'equidistant':
if target == 3:
self.failUnlessEqual(chosenchunks, [0, 3, 7])
elif strategy == 'random':
# none is selected twice
self.failUnless(len(set(chosenchunks)) == len(chosenchunks))
self.failUnless(target == len(chosenchunks))
else:
raise RuntimeError, "Add unittest for strategy %s" \
% strategy
def test_svms(self, clf):
knows_probabilities = \
'probabilities' in clf.ca.keys() and clf.params.probability
enable_ca = ['estimates']
if knows_probabilities:
enable_ca += ['probabilities']
clf.ca.change_temporarily(enable_ca = enable_ca)
spl = Splitter('train', count=2)
traindata, testdata = list(spl.generate(datasets['uni2small']))
clf.train(traindata)
predicts = clf.predict(testdata.samples)
# values should be different from predictions for SVMs we have
self.failUnless(np.any(predicts != clf.ca.estimates))
if knows_probabilities and clf.ca.is_set('probabilities'):
# XXX test more thoroughly what we are getting here ;-)
self.failUnlessEqual( len(clf.ca.probabilities),
len(testdata.samples) )
clf.ca.reset_changed_temporarily()
def _forward_dataset(self, ds):
if self.__chunks_attr is None:
return self._forward_dataset_helper(ds)
else:
# strip down dataset to speedup local processing
if self.__attr_strategy == 'remove':
keep_sa = []
else:
keep_sa = None
proc_ds = ds.copy(deep=False, sa=keep_sa, fa=[], a=[])
# process all chunks individually
# use a customsplitter to speed-up splitting
spl = Splitter(self.__chunks_attr)
dses = [self._forward_dataset_helper(d)
for d in spl.generate(proc_ds)]
# and merge them again
mds = vstack(dses)
# put back attributes
mds.fa.update(ds.fa)
mds.a.update(ds.a)
return mds
def test_n_group_split(self):
"""Test NGroupSplitter alongside with the reversal of the
order of spit out datasets
"""
# Test 2 groups like HalfSplitter first
hs = NGroupPartitioner(2)
for isreversed, splitter in enumerate((hs, hs)):
if isreversed:
spl = Splitter(attr='partitions', reverse=True)
else:
spl = Splitter(attr='partitions')
splits = [ list(spl.generate(p)) for p in hs.generate(self.data) ]
self.failUnless(len(splits) == 2)
for i, p in enumerate(splits):
self.failUnless( len(p) == 2 )
self.failUnless( p[0].nsamples == 50 )
self.failUnless( p[1].nsamples == 50 )
assert_array_equal(splits[0][1-isreversed].sa['chunks'].unique,
[0, 1, 2, 3, 4])
assert_array_equal(splits[0][isreversed].sa['chunks'].unique,
[5, 6, 7, 8, 9])
assert_array_equal(splits[1][1-isreversed].sa['chunks'].unique,
[5, 6, 7, 8, 9])
assert_array_equal(splits[1][isreversed].sa['chunks'].unique,
[0, 1, 2, 3, 4])
# check if it works on pure odd and even chunk ids
moresplits = [ list(spl.generate(p)) for p in hs.generate(splits[0][0])]
for split in moresplits:
self.failUnless(split[0] != None)
self.failUnless(split[1] != None)
# now test more groups
s5 = NGroupPartitioner(5)
# get the splits
for isreversed, s5splitter in enumerate((s5, s5)):
if isreversed:
spl = Splitter(attr='partitions', reverse=True)
else:
spl = Splitter(attr='partitions')
splits = [ list(spl.generate(p)) for p in s5splitter.generate(self.data) ]
# must have 10 splits
self.failUnless(len(splits) == 5)
# check split content
assert_array_equal(splits[0][1-isreversed].sa['chunks'].unique,
[0, 1])
assert_array_equal(splits[0][isreversed].sa['chunks'].unique,
[2, 3, 4, 5, 6, 7, 8, 9])
assert_array_equal(splits[1][1-isreversed].sa['chunks'].unique,
[2, 3])
assert_array_equal(splits[1][isreversed].sa['chunks'].unique,
[0, 1, 4, 5, 6, 7, 8, 9])
# ...
assert_array_equal(splits[4][1-isreversed].sa['chunks'].unique,
[8, 9])
assert_array_equal(splits[4][isreversed].sa['chunks'].unique,
[0, 1, 2, 3, 4, 5, 6, 7])
# Test for too many groups
def splitcall(spl, dat):
return list(spl.generate(dat))
s20 = NGroupPartitioner(20)
self.assertRaises(ValueError,splitcall,s20,self.data)
def _sl_call(self, dataset, roi_ids, nproc):
"""Call to GNBSearchlight
"""
# Local bindings
gnb = self.gnb
params = gnb.params
generator = self.generator
errorfx = self.errorfx
qe = self.queryengine
## if False:
## class A(Learner):
## pass
## self = A()
## import numpy as np
## from mvpa.clfs.gnb import GNB
## from mvpa.generators.partition import NFoldPartitioner
## from mvpa.misc.errorfx import mean_mismatch_error
## from mvpa.testing.datasets import datasets as tdatasets
## from mvpa.datasets import Dataset
## from mvpa.misc.neighborhood import IndexQueryEngine, Sphere
## from mvpa.clfs.distance import absmin_distance
## import time
## if __debug__:
## from mvpa.base import debug
## debug.active += ['SLC.*']
## # XXX is it that ugly?
## debug.active.pop(debug.active.index('SLC_'))
## debug.metrics += ['reltime']
## dataset = tdatasets['3dlarge'].copy()
## dataset.fa['voxel_indices'] = dataset.fa.myspace
## sphere = Sphere(radius=1,
## distance_func=absmin_distance)
## qe = IndexQueryEngine(myspace=sphere)
## # Fracisco's data
## #dataset = ds_fp
## qe = IndexQueryEngine(voxel_indices=sphere)
## qe.train(dataset)
## roi_ids = np.arange(dataset.nfeatures)
## gnb = GNB()
## params = gnb.params
## generator = NFoldPartitioner()
## errorfx = mean_mismatch_error
if __debug__:
time_start = time.time()
targets_sa_name = gnb.get_space()
targets_sa = dataset.sa[targets_sa_name]
if __debug__:
debug_slc_ = 'SLC_' in debug.active
# get the dataset information into easy vars
X = dataset.samples
if len(X.shape) != 2:
raise ValueError, \
'Unlike GNB, GNBSearchlight (for now) operates on already' \
'flattened datasets'
labels = targets_sa.value
ulabels = targets_sa.unique
nlabels = len(ulabels)
label2index = dict((l, il) for il, l in enumerate(ulabels))
labels_numeric = np.array([label2index[l] for l in labels])
ulabels_numeric = [label2index[l] for l in ulabels]
# set the feature dimensions
nsamples = len(X)
nrois = len(roi_ids)
s_shape = X.shape[1:] # shape of a single sample
# The shape of results
r_shape = (nrois,) + X.shape[2:]
#
# Everything toward optimization ;)
#
# Silly Yarik thinks that it might be worth to pre-compute
# statistics per each feature within a block of the samples
# which always come together in splits -- most often it is a
# (chunk, label) combination, but since we simply use a
# generator -- who knows! Therefore lets figure out what are
# those blocks and operate on them instead of original samples.
#
# After additional thinking about this -- probably it would be
# just minor additional improvements (ie not worth it) but
# since it is coded already -- let it be so
# 1. Query generator for the splits we will have
if __debug__:
debug('SLC',
'Phase 1. Initializing partitions using %s on %s'
% (generator, dataset))
# Lets just create a dummy ds which will store for us actual sample
# indicies
# XXX we could make it even more lightweight I guess...
dataset_indicies = Dataset(np.arange(nsamples), sa=dataset.sa)
splitter = Splitter(attr=generator.get_space())
splits = list(tuple(splitter.generate(ds_))
for ds_ in generator.generate(dataset_indicies))
nsplits = len(splits)
# 2. Figure out the new 'chunks x labels' blocks of combinations
# of samples
if __debug__:
debug('SLC',
'Phase 2. Blocking data for %i splits and %i labels'
% (nsplits, nlabels))
# array of indicies for label, split1, split2, ...
# through which we will pass later on to figure out
# unique combinations
combinations = np.ones((nsamples, 1+nsplits), dtype=int)*-1
# labels
combinations[:, 0] = labels_numeric
for ipartition, (split1, split2) in enumerate(splits):
combinations[split1.samples[:, 0], 1+ipartition] = 1
combinations[split2.samples[:, 0], 1+ipartition] = 2
# Check for over-sampling, i.e. no same sample used twice here
if not (len(np.unique(split1.samples[:, 0])) == len(split1) and
len(np.unique(split2.samples[:, 0])) == len(split2)):
raise RuntimeError(
"GNBSearchlight needs a partitioner which does not reuse "
"the same the same samples more than once")
# sample descriptions -- should be unique for
# samples within the same block
descriptions = [tuple(c) for c in combinations]
udescriptions = sorted(list(set(descriptions)))
nblocks = len(udescriptions)
description2block = dict([(d, i) for i, d in enumerate(udescriptions)])
# Indices for samples to point to their block
sample2block = np.array([description2block[d] for d in descriptions])
# 3. Compute statistics per each block
#
if __debug__:
debug('SLC',
'Phase 3. Computing statistics for %i blocks' % (nblocks,))
#
# reusable containers which should stay of the same size
#
# sums and sums of squares per each block
sums = np.zeros((nblocks, ) + s_shape)
# sums of squares
sums2 = np.zeros((nblocks, ) + s_shape)
# per each label:
means = np.zeros((nlabels, ) + s_shape)
# means of squares for stddev computation
means2 = np.zeros((nlabels, ) + s_shape)
variances = np.zeros((nlabels, ) + s_shape)
# degenerate dimension are added for easy broadcasting later on
nsamples_per_class = np.zeros((nlabels,) + (1,)*len(s_shape))
# results
results = np.zeros((nsplits,) + r_shape)
block_counts = np.zeros((nblocks,))
block_labels = [None] * nblocks
X2 = np.square(X)
# silly way for now
for l, s, s2, ib in zip(labels_numeric, X, X2, sample2block):
sums[ib] += s
sums2[ib] += s2
block_counts[ib] += 1
if block_labels[ib] is None:
block_labels[ib] = l
else:
assert(block_labels[ib] == l)
block_labels = np.asanyarray(block_labels)
# additional silly tests for paranoid
assert(block_labels.dtype.kind is 'i')
# 4. Lets deduce all neighbors... might need to be RF into the
# parallel part later on
if __debug__:
debug('SLC',
'Phase 4. Deducing neighbors information for %i ROIs'
% (nrois,))
roi_fids = [qe.query_byid(f) for f in roi_ids]
nroi_fids = len(roi_fids)
# makes sense to waste precious ms only if ca is enabled
if self.ca.is_enabled('roi_sizes'):
roi_sizes = [len(x) for x in roi_fids]
else:
roi_sizes = []
indexsum = self._indexsum
if indexsum == 'sparse':
if __debug__:
debug('SLC',
'Phase 4b. Converting neighbors to sparse matrix '
'representation')
# convert to "sparse representation" where column j contains
# 1s only at the roi_fids[j] indices
roi_fids = inds_to_coo(roi_fids,
shape=(dataset.nfeatures, nroi_fids))
indexsum_fx = lastdim_columnsums_spmatrix
elif indexsum == 'fancy':
indexsum_fx = lastdim_columnsums_fancy_indexing
else:
raise ValueError, \
"Do not know how to deal with indexsum=%s" % indexsum
# 5. Lets do actual "splitting" and "classification"
if __debug__:
debug('SLC', 'Phase 5. Major loop' )
for isplit, split in enumerate(splits):
if __debug__:
debug('SLC', ' Split %i out of %i' % (isplit, nsplits))
# figure out for a given splits the blocks we want to work
# with
# sample_indicies
training_sis = split[0].samples[:, 0]
# convert to blocks training split
training_bis = np.unique(sample2block[training_sis])
# now lets do our GNB business
training_nsamples = 0
for il, l in enumerate(ulabels_numeric):
bis_il = training_bis[block_labels[training_bis] == l]
nsamples_per_class[il] = N_float = \
float(np.sum(block_counts[bis_il]))
training_nsamples += N_float
if N_float == 0.0:
variances[il] = means[il] = means2[il] = 0.
else:
means[il] = np.sum(sums[bis_il], axis=0) / N_float
# Not yet normed
means2[il] = np.sum(sums2[bis_il], axis=0)
## Actually compute the non-0 variances
non0labels = (nsamples_per_class.squeeze() != 0)
if np.all(non0labels):
# For a possible tiny speed up avoiding copying and
# using (no) slicing
non0labels = slice(None)
if params.common_variance:
variances[:] = \
np.sum(means2 - nsamples_per_class*np.square(means),
axis=0) \
/ training_nsamples
else:
variances[non0labels] = \
(means2 - nsamples_per_class*np.square(means))[non0labels] \
/ nsamples_per_class[non0labels]
# assign priors
priors = gnb._get_priors(
nlabels, training_nsamples, nsamples_per_class)
# proceed in a way we have in GNB code with logprob=True,
# i.e. operating within the exponents -- should lead to some
# performance advantage
norm_weight = -0.5 * np.log(2*np.pi*variances)
# last added dimension would be for ROIs
logpriors = np.log(priors[:, np.newaxis, np.newaxis])
if __debug__:
debug('SLC', " 'Training' is done")
# Now it is time to "classify" our samples.
# and for that we first need to compute corresponding
# probabilities (or may be un
data = X[split[1].samples[:, 0]]
targets = labels_numeric[split[1].samples[:, 0]]
# argument of exponentiation
scaled_distances = \
-0.5 * (((data - means[:, np.newaxis, ...])**2) \
/ variances[:, np.newaxis, ...])
# incorporate the normalization from normals
lprob_csfs = norm_weight[:, np.newaxis, ...] + scaled_distances
## First we need to reshape to get class x samples x features
lprob_csf = lprob_csfs.reshape(lprob_csfs.shape[:2] + (-1,))
## Now we come to naive part which requires looping
## through all spheres
if __debug__:
debug('SLC', " Doing 'Searchlight'")
# resultant logprobs for each class x sample x roi
lprob_cs_sl = np.zeros(lprob_csfs.shape[:2] + (nroi_fids,))
indexsum_fx(lprob_csf, roi_fids, out=lprob_cs_sl)
lprob_cs_sl += logpriors
lprob_cs_cp_sl = lprob_cs_sl
# for each of the ROIs take the class with maximal (log)probability
predictions = lprob_cs_cp_sl.argmax(axis=0)
# no need to map back [self.ulabels[c] for c in winners]
#predictions = winners
# assess the errors
if __debug__:
debug('SLC', " Assessing accuracies")
if errorfx is mean_mismatch_error:
results[isplit, :] = \
(predictions != targets[:, None]).sum(axis=0) \
/ float(len(targets))
else:
# somewhat silly but a way which allows to use pre-crafted
# error functions without a chance to screw up
for i, fpredictions in enumerate(predictions.T):
results[isplit, i] = errorfx(fpredictions, targets)
if __debug__:
debug('SLC', "GNBSearchlight is done in %.3g sec" %
(time.time() - time_start))
return Dataset(results), roi_sizes