def get_cluster_sizes(ds, cluster_counter=None):
"""Computer cluster sizes from all samples in a boolean dataset.
Individually for each sample, in the input dataset, clusters of non-zero
values will be determined after reverse-applying any transformation of the
dataset's mapper (if any).
Parameters
----------
ds : dataset or array
A dataset with boolean samples.
cluster_counter : list or None
If not None, the given list is extended with the cluster sizes computed
from the present input dataset. Otherwise, a new list is generated.
Returns
-------
list
Unsorted list of cluster sizes from all samples in the input dataset
(optionally appended to any values passed via ``cluster_counter``).
"""
# XXX input needs to be boolean for the cluster size calculation to work
if cluster_counter is None:
cluster_counter = Counter()
mapper = IdentityMapper()
data = np.asanyarray(ds)
if hasattr(ds, 'a') and 'mapper' in ds.a:
mapper = ds.a.mapper
for i in xrange(len(ds)):
osamp = mapper.reverse1(data[i])
m_clusters = _get_map_cluster_sizes(osamp)
cluster_counter.update(m_clusters)
return cluster_counter
def test_identity_mapper(s):
idm = IdentityMapper()
# doesn't matter what you throw at it
assert_true(idm.forward(s) is s)
assert_true(idm.forward1(s) is s)
assert_true(idm.reverse(s) is s)
assert_true(idm.reverse1(s) is s)
def test_identity_mapper(s):
idm = IdentityMapper()
# doesn't matter what you throw at it
assert_true(idm.forward(s) is s)
assert_true(idm.forward1(s) is s)
assert_true(idm.reverse(s) is s)
assert_true(idm.reverse1(s) is s)
def test_identity_mapper(s):
idm = IdentityMapper()
# doesn't matter what you throw at it
assert_true(idm.forward(s) is s)
assert_true(idm.forward1(s) is s)
assert_true(idm.reverse(s) is s)
assert_true(idm.reverse1(s) is s)
# even like this it should work, but type conversion
# can happen
assert_array_equal(_verified_reverse1(idm, s), s)
assert_array_equal(idm.reverse1(s), s)
def _call(self, ds):
if len(ds) > 1:
# average all samples into one, assuming we got something like one
# sample per subject as input
avgr = mean_sample()
ds = avgr(ds)
# threshold input; at this point we only have one sample left
thrd = ds.samples[0] > self._thrmap
# mapper default
mapper = IdentityMapper()
# overwrite if possible
if hasattr(ds, 'a') and 'mapper' in ds.a:
mapper = ds.a.mapper
# reverse-map input
othrd = _verified_reverse1(mapper, thrd)
# TODO: what is your purpose in life osamp? ;-)
osamp = _verified_reverse1(mapper, ds.samples[0])
# prep output dataset
outds = ds.copy(deep=False)
outds.fa['featurewise_thresh'] = self._thrmap
# determine clusters
labels, num = measurements.label(othrd,structure=np.ones([3,3,3]))
area = measurements.sum(othrd,
labels,
index=np.arange(1, num + 1)).astype(int)
com = measurements.center_of_mass(
osamp, labels=labels, index=np.arange(1, num + 1))
maxpos = measurements.maximum_position(
osamp, labels=labels, index=np.arange(1, num + 1))
# for the rest we need the labels flattened
labels = mapper.forward1(labels)
# relabel clusters starting with the biggest and increase index with
# decreasing size
ordered_labels = np.zeros(labels.shape, dtype=int)
ordered_area = np.zeros(area.shape, dtype=int)
ordered_com = np.zeros((num, len(osamp.shape)), dtype=float)
ordered_maxpos = np.zeros((num, len(osamp.shape)), dtype=float)
for i, idx in enumerate(np.argsort(area)):
ordered_labels[labels == idx + 1] = num - i
# kinda ugly, but we are looping anyway
ordered_area[i] = area[idx]
ordered_com[i] = com[idx]
ordered_maxpos[i] = maxpos[idx]
labels = ordered_labels
area = ordered_area[::-1]
com = ordered_com[::-1]
maxpos = ordered_maxpos[::-1]
del ordered_labels # this one can be big
# store cluster labels after forward-mapping
outds.fa['clusters_featurewise_thresh'] = labels.copy()
# location info
outds.a['clusterlocations'] = \
np.rec.fromarrays(
[com, maxpos], names=('center_of_mass', 'max'))
# update cluster size histogram with the actual result to get a
# proper lower bound for p-values
# this will make a copy, because the original matrix is int
cluster_probs_raw = _transform_to_pvals(
area, self._null_cluster_sizes.astype('float'))
clusterstats = (
[area, cluster_probs_raw],
['size', 'prob_raw']
)
# evaluate a bunch of stats for all clusters
morestats = {}
for cid in xrange(len(area)):
# keep clusters on outer loop, because selection is more expensive
clvals = ds.samples[0, labels == cid + 1]
for id_, fx in (
('mean', np.mean),
('median', np.median),
('min', np.min),
('max', np.max),
('std', np.std)):
stats = morestats.get(id_, [])
stats.append(fx(clvals))
morestats[id_] = stats
for k, v in morestats.items():
clusterstats[0].append(v)
clusterstats[1].append(k)
if self.params.multicomp_correction is not None:
# do a local import as only this tiny portion needs statsmodels
import statsmodels.stats.multitest as smm
rej, probs_corr = smm.multipletests(
cluster_probs_raw,
alpha=self.params.fwe_rate,
method=self.params.multicomp_correction)[:2]
# store corrected per-cluster probabilities
clusterstats[0].append(probs_corr)
clusterstats[1].append('prob_corrected')
# remove cluster labels that did not pass the FWE threshold
for i, r in enumerate(rej):
if not r:
labels[labels == i + 1] = 0
outds.fa['clusters_fwe_thresh'] = labels
outds.a['clusterstats'] = \
np.rec.fromarrays(clusterstats[0], names=clusterstats[1])
return outds
def nohyper(dss):
return [IdentityMapper() for ds in dss]
def timesegments_classification(dss,
hyper=None,
part1=HalfPartitioner(),
part2=NFoldPartitioner(attr='subjects'),
window_size=6,
overlapping_windows=True,
distance='correlation',
do_zscore=True):
"""Time-segment classification across subjects using Hyperalignment
Parameters
----------
dss : list of datasets
Datasets to benchmark on. Usually a single dataset per subject.
hyper : Hyperalignment-like, optional
Beast which if called on a list of datasets should spit out trained
mappers. If not specified, `IdentityMapper`s will be used
part1 : Partitioner, optional
Partitioner to split data for hyperalignment "cross-validation"
part2 : Partitioner, optional
Partitioner for CV within the hyperalignment test split
window_size : int, optional
How many temporal points to consider for a classification sample
overlapping_windows : bool, optional
Strategy to how create and classify "samples" for classification. If
True -- `window_size` samples from each time point (but trailing ones)
constitute a sample, and upon "predict" `window_size` of samples around
each test point is not considered. If False -- samples are just taken
(with training and testing splits) at `window_size` step from one to
another.
do_zscore : bool, optional
Perform zscoring (overall, not per-chunk) for each dataset upon
partitioning with part1
...
"""
# Generate outer-most partitioning ()
parts = [copy.deepcopy(part1).generate(ds) for ds in dss]
iter = 1
errors = []
while True:
try:
dss_partitioned = [p.next() for p in parts]
except StopIteration:
# we are done -- no more partitions
break
if __debug__:
debug("BM", "Iteration %d", iter)
dss_train, dss_test = zip(*[
list(Splitter("partitions").generate(ds)) for ds in dss_partitioned
])
# TODO: allow for doing feature selection
if do_zscore:
for ds in dss_train + dss_test:
zscore(ds, chunks_attr=None)
if hyper is not None:
# since otherwise it would remember previous loop dataset as the "commonspace"
# Now let's do hyperalignment but on a copy in each loop iteration
hyper_ = copy.deepcopy(hyper)
mappers = hyper_(dss_train)
else:
mappers = [IdentityMapper() for ds in dss_train]
dss_test_aligned = [
mapper.forward(ds) for mapper, ds in zip(mappers, dss_test)
]
# assign .sa.subjects to those datasets
for i, ds in enumerate(dss_test_aligned):
# part2.attr is by default "subjects"
ds.sa[part2.attr] = [i]
dss_test_bc = []
for ds in dss_test_aligned:
if overlapping_windows:
startpoints = range(len(ds) - window_size + 1)
else:
startpoints = _get_nonoverlapping_startpoints(
len(ds), window_size)
bm = BoxcarMapper(startpoints, window_size)
bm.train(ds)
ds_ = bm.forward(ds)
ds_.sa['startpoints'] = startpoints
# reassign subjects so they are not arrays
def assign_unique(ds, sa):
ds.sa[sa] = [
np.asscalar(np.unique(x)) for x in ds.sa[sa].value
]
assign_unique(ds_, part2.attr)
fm = FlattenMapper()
fm.train(ds_)
dss_test_bc.append(ds_.get_mapped(fm))
ds_test = vstack(dss_test_bc)
# Perform classification across subjects comparing against mean
# spatio-temporal pattern of other subjects
errors_across_subjects = []
for ds_test_part in part2.generate(ds_test):
ds_train_, ds_test_ = list(
Splitter("partitions").generate(ds_test_part))
# average across subjects to get a representative pattern per timepoint
ds_train_ = mean_group_sample(['startpoints'])(ds_train_)
assert (ds_train_.shape == ds_test_.shape)
if distance == 'correlation':
# TODO: redo more efficiently since now we are creating full
# corrcoef matrix. Also we might better just take a name for
# the pdist measure but then implement them efficiently
# (i.e. without hstacking both pieces together first)
dist = 1 - np.corrcoef(
ds_train_, ds_test_)[len(ds_test_):, :len(ds_test_)]
else:
raise NotImplementedError
if overlapping_windows:
dist = wipe_out_offdiag(dist, window_size)
winners = np.argmin(dist, axis=1)
error = np.mean(winners != np.arange(len(winners)))
errors_across_subjects.append(error)
errors.append(errors_across_subjects)
iter += 1
errors = np.array(errors)
if __debug__:
debug(
"BM", "Finished with %s array of errors. Mean error %.2f" %
(errors.shape, np.mean(errors)))
return errors
def _call(self, ds):
if len(ds) > 1:
# average all samples into one, assuming we got something like one
# sample per subject as input
avgr = mean_sample()
ds = avgr(ds)
# threshold input; at this point we only have one sample left
thrd = ds.samples[0] > self._thrmap
# mapper default
mapper = IdentityMapper()
# overwrite if possible
if hasattr(ds, 'a') and 'mapper' in ds.a:
mapper = ds.a.mapper
# reverse-map input
osamp = mapper.reverse1(thrd)
# prep output dataset
outds = ds.copy(deep=False)
outds.fa['featurewise_thresh'] = self._thrmap
# determine clusters
labels, num = measurements.label(osamp)
area = measurements.sum(osamp,
labels,
index=np.arange(1, num + 1)).astype(int)
# for the rest we need the labels flattened
labels = mapper.forward1(labels)
# relabel clusters starting with the biggest and increase index with
# decreasing size
ordered_labels = np.zeros(labels.shape, dtype=int)
ordered_area = np.zeros(area.shape, dtype=int)
for i, idx in enumerate(np.argsort(area)):
ordered_labels[labels == idx + 1] = num - i
ordered_area[i] = area[idx]
area = ordered_area[::-1]
labels = ordered_labels
del ordered_labels # this one can be big
# store cluster labels after forward-mapping
outds.fa['clusters_featurewise_thresh'] = labels.copy()
# update cluster size histogram with the actual result to get a
# proper lower bound for p-values
# this will make a copy, because the original matrix is int
cluster_probs_raw = _transform_to_pvals(
area, self._null_cluster_sizes.astype('float'))
if self.params.multicomp_correction is None:
probs_corr = np.array(cluster_probs_raw)
rej = probs_corr <= self.params.fwe_rate
outds.a['clusterstats'] = \
np.rec.fromarrays(
[area, cluster_probs_raw], names=('size', 'prob_raw'))
else:
# do a local import as only this tiny portion needs statsmodels
import statsmodels.stats.multitest as smm
rej, probs_corr = smm.multipletests(
cluster_probs_raw,
alpha=self.params.fwe_rate,
method=self.params.multicomp_correction)[:2]
# store corrected per-cluster probabilities
outds.a['clusterstats'] = \
np.rec.fromarrays(
[area, cluster_probs_raw, probs_corr],
names=('size', 'prob_raw', 'prob_corrected'))
# remove cluster labels that did not pass the FWE threshold
for i, r in enumerate(rej):
if not r:
labels[labels == i + 1] = 0
outds.fa['clusters_fwe_thresh'] = labels
return outds
