def _test_compare_to_old(self):
"""Good just to compare if I didn't screw up anything... treat
it as a regression test
"""
import mvpa2.mappers.wavelet_ as wavelet_
ds = datasets['uni2medium']
d2d = ds.samples
ws = 16 # size of timeline for wavelet
sp = np.arange(ds.nsamples-ws*2) + ws
# create 3D instance (samples x timepoints x channels)
bcm = BoxcarMapper(sp, ws)
d3d = bcm.forward(d2d)
# use wavelet mapper
for wdm, wdm_ in ((WaveletTransformationMapper(),
wavelet_.WaveletTransformationMapper()),
(WaveletPacketMapper(),
wavelet_.WaveletPacketMapper()),):
d3d_wd = wdm(d3d)
d3d_wd_ = wdm_(d3d)
self.failUnless((d3d_wd == d3d_wd_).all(),
msg="We should have got same result with old and new code. "
"Got %s and %s" % (d3d_wd, d3d_wd_))
def _test_compare_to_old(self):
"""Good just to compare if I didn't screw up anything... treat
it as a regression test
"""
import mvpa2.mappers.wavelet_ as wavelet_
ds = datasets['uni2medium']
d2d = ds.samples
ws = 16 # size of timeline for wavelet
sp = np.arange(ds.nsamples-ws*2) + ws
# create 3D instance (samples x timepoints x channels)
bcm = BoxcarMapper(sp, ws)
d3d = bcm.forward(d2d)
# use wavelet mapper
for wdm, wdm_ in ((WaveletTransformationMapper(),
wavelet_.WaveletTransformationMapper()),
(WaveletPacketMapper(),
wavelet_.WaveletPacketMapper()),):
d3d_wd = wdm(d3d)
d3d_wd_ = wdm_(d3d)
self.assertTrue((d3d_wd == d3d_wd_).all(),
msg="We should have got same result with old and new code. "
"Got %s and %s" % (d3d_wd, d3d_wd_))
def test_simple_wdm(self):
"""
"""
ds = datasets['uni2medium']
d2d = ds.samples
ws = 15 # size of timeline for wavelet
sp = np.arange(ds.nsamples - ws * 2) + ws
# create 3D instance (samples x timepoints x channels)
bcm = BoxcarMapper(sp, ws)
d3d = bcm.forward(d2d)
# use wavelet mapper
wdm = WaveletTransformationMapper()
d3d_wd = wdm.forward(d3d)
d3d_swap = d3d.swapaxes(1, 2)
self.assertRaises(ValueError,
WaveletTransformationMapper,
wavelet='bogus')
self.assertRaises(ValueError,
WaveletTransformationMapper,
mode='bogus')
# use wavelet mapper
for wdm, wdm_swap in ((WaveletTransformationMapper(),
WaveletTransformationMapper(dim=2)),
(WaveletPacketMapper(),
WaveletPacketMapper(dim=2))):
for dd, dd_swap in ((d3d, d3d_swap), (d2d, None)):
dd_wd = wdm.forward(dd)
if dd_swap is not None:
dd_wd_swap = wdm_swap.forward(dd_swap)
self.assertTrue(
(dd_wd == dd_wd_swap.swapaxes(1, 2)).all(),
msg="We should have got same result with swapped "
"dimensions and explicit mentioining of it. "
"Got %s and %s" % (dd_wd, dd_wd_swap))
# some sanity checks
self.assertTrue(dd_wd.shape[0] == dd.shape[0])
if not isinstance(wdm, WaveletPacketMapper):
# we can do reverse only for DWT
dd_rev = wdm.reverse(dd_wd)
# inverse transform might be not exactly as the
# input... but should be very close ;-)
self.assertEqual(dd_rev.shape,
dd.shape,
msg="Shape should be the same after iDWT")
diff = np.linalg.norm(dd - dd_rev)
ornorm = np.linalg.norm(dd)
self.assertTrue(diff / ornorm < 1e-10)
def test_simple_wdm(self):
"""
"""
ds = datasets['uni2medium']
d2d = ds.samples
ws = 15 # size of timeline for wavelet
sp = np.arange(ds.nsamples-ws*2) + ws
# create 3D instance (samples x timepoints x channels)
bcm = BoxcarMapper(sp, ws)
d3d = bcm.forward(d2d)
# use wavelet mapper
wdm = WaveletTransformationMapper()
d3d_wd = wdm.forward(d3d)
d3d_swap = d3d.swapaxes(1,2)
self.failUnlessRaises(ValueError, WaveletTransformationMapper,
wavelet='bogus')
self.failUnlessRaises(ValueError, WaveletTransformationMapper,
mode='bogus')
# use wavelet mapper
for wdm, wdm_swap in ((WaveletTransformationMapper(),
WaveletTransformationMapper(dim=2)),
(WaveletPacketMapper(),
WaveletPacketMapper(dim=2))):
for dd, dd_swap in ((d3d, d3d_swap),
(d2d, None)):
dd_wd = wdm.forward(dd)
if dd_swap is not None:
dd_wd_swap = wdm_swap.forward(dd_swap)
self.failUnless((dd_wd == dd_wd_swap.swapaxes(1,2)).all(),
msg="We should have got same result with swapped "
"dimensions and explicit mentioining of it. "
"Got %s and %s" % (dd_wd, dd_wd_swap))
# some sanity checks
self.failUnless(dd_wd.shape[0] == dd.shape[0])
if not isinstance(wdm, WaveletPacketMapper):
# we can do reverse only for DWT
dd_rev = wdm.reverse(dd_wd)
# inverse transform might be not exactly as the
# input... but should be very close ;-)
self.failUnlessEqual(dd_rev.shape, dd.shape,
msg="Shape should be the same after iDWT")
diff = np.linalg.norm(dd - dd_rev)
ornorm = np.linalg.norm(dd)
self.failUnless(diff/ornorm < 1e-10)
def test_simple_wp1_level(self):
"""
"""
ds = datasets['uni2large']
d2d = ds.samples
ws = 50 # size of timeline for wavelet
sp = (np.arange(ds.nsamples - ws * 2) + ws)[:4]
# create 3D instance (samples x timepoints x channels)
bcm = BoxcarMapper(sp, ws)
d3d = bcm.forward(d2d)
# use wavelet mapper
wdm = WaveletPacketMapper(level=2, wavelet='sym2')
d3d_wd = wdm.forward(d3d)
# Check dimensionality
d3d_wds, d3ds = d3d_wd.shape, d3d.shape
self.assertTrue(len(d3d_wds) == len(d3ds) + 1)
self.assertTrue(d3d_wds[1] * d3d_wds[2] >= d3ds[1])
self.assertTrue(d3d_wds[0] == d3ds[0])
self.assertTrue(d3d_wds[-1] == d3ds[-1])
#print d2d.shape, d3d.shape, d3d_wd.shape
if externals.exists('pywt wp reconstruct'):
# Test reverse -- should be identical
# we can do reverse only for DWT
d3d_rev = wdm.reverse(d3d_wd)
# inverse transform might be not exactly as the
# input... but should be very close ;-)
self.assertEqual(d3d_rev.shape,
d3d.shape,
msg="Shape should be the same after iDWT")
diff = np.linalg.norm(d3d - d3d_rev)
ornorm = np.linalg.norm(d3d)
skip_if_no_external('pywt wp reconstruct fixed')
self.assertTrue(diff / ornorm < 1e-10)
else:
self.assertRaises(NotImplementedError, wdm.reverse, d3d_wd)
def test_simple_wp1_level(self):
"""
"""
ds = datasets['uni2large']
d2d = ds.samples
ws = 50 # size of timeline for wavelet
sp = (np.arange(ds.nsamples - ws*2) + ws)[:4]
# create 3D instance (samples x timepoints x channels)
bcm = BoxcarMapper(sp, ws)
d3d = bcm.forward(d2d)
# use wavelet mapper
wdm = WaveletPacketMapper(level=2, wavelet='sym2')
d3d_wd = wdm.forward(d3d)
# Check dimensionality
d3d_wds, d3ds = d3d_wd.shape, d3d.shape
self.failUnless(len(d3d_wds) == len(d3ds)+1)
self.failUnless(d3d_wds[1] * d3d_wds[2] >= d3ds[1])
self.failUnless(d3d_wds[0] == d3ds[0])
self.failUnless(d3d_wds[-1] == d3ds[-1])
#print d2d.shape, d3d.shape, d3d_wd.shape
if externals.exists('pywt wp reconstruct'):
# Test reverse -- should be identical
# we can do reverse only for DWT
d3d_rev = wdm.reverse(d3d_wd)
# inverse transform might be not exactly as the
# input... but should be very close ;-)
self.failUnlessEqual(d3d_rev.shape, d3d.shape,
msg="Shape should be the same after iDWT")
diff = np.linalg.norm(d3d - d3d_rev)
ornorm = np.linalg.norm(d3d)
skip_if_no_external('pywt wp reconstruct fixed')
self.failUnless(diff/ornorm < 1e-10)
else:
self.failUnlessRaises(NotImplementedError, wdm.reverse, d3d_wd)
def test_simpleboxcar():
data = np.atleast_2d(np.arange(10)).T
sp = np.arange(10)
# check if stupid thing don't work
assert_raises(ValueError, BoxcarMapper, sp, 0)
# now do an identity transformation
bcm = BoxcarMapper(sp, 1)
trans = bcm.forward(data)
# ,0 is a feature below, so we get explicit 2D out of 1D
assert_array_equal(trans[:, 0], data)
# now check for illegal boxes
if __debug__:
# condition is checked only in __debug__
assert_raises(ValueError, BoxcarMapper(sp, 2).train, data)
# now something that should work
nbox = 9
boxlength = 2
sp = np.arange(nbox)
bcm = BoxcarMapper(sp, boxlength)
trans = bcm.forward(data)
# check that is properly upcasts the dimensionality
assert_equal(trans.shape, (nbox, boxlength) + data.shape[1:])
# check actual values, squeezing the last dim for simplicity
assert_array_equal(trans.squeeze(),
np.vstack((np.arange(9), np.arange(9) + 1)).T)
# now test for proper data shape
data = np.ones((10, 3, 4, 2))
sp = [2, 4, 3, 5]
trans = BoxcarMapper(sp, 4).forward(data)
assert_equal(trans.shape, (4, 4, 3, 4, 2))
# test reverse
data = np.arange(240).reshape(10, 3, 4, 2)
sp = [2, 4, 3, 5]
boxlength = 2
m = BoxcarMapper(sp, boxlength)
m.train(data)
mp = m.forward(data)
assert_equal(mp.shape, (4, 2, 3, 4, 2))
# try full reconstruct
mr = m.reverse(mp)
# shape has to match
assert_equal(mr.shape, (len(sp) * boxlength, ) + data.shape[1:])
# only known samples are part of the results
assert_true((mr >= 24).all())
assert_true((mr < 168).all())
# check proper reconstruction of non-conflicting sample
assert_array_equal(mr[0].ravel(), np.arange(48, 72))
# check proper reconstruction of samples being part of multiple
# mapped samples
assert_array_equal(mr[1].ravel(), np.arange(72, 96))
# test reverse of a single sample
singlesample = np.arange(48).reshape(2, 3, 4, 2)
assert_array_equal(singlesample, m.reverse1(singlesample))
# now in a dataset
ds = Dataset([singlesample])
assert_equal(ds.shape, (1, ) + singlesample.shape)
# after reverse mapping the 'sample axis' should vanish and the original 3d
# shape of the samples should be restored
assert_equal(ds.shape[1:], m.reverse(ds).shape)
# multiple samples should just be concatenated along the samples axis
ds = Dataset([singlesample, singlesample])
assert_equal((np.prod(ds.shape[:2]), ) + singlesample.shape[1:],
m.reverse(ds).shape)
# should not work for shape mismatch, but it does work and is useful when
# reverse mapping sample attributes
#assert_raises(ValueError, m.reverse, singlesample[0])
# check broadcasting of 'raw' samples into proper boxcars on forward()
bc = m.forward1(np.arange(24).reshape(3, 4, 2))
assert_array_equal(bc, np.array(2 * [np.arange(24).reshape(3, 4, 2)]))
def test_datasetmapping():
# 6 samples, 4X2 features
data = np.arange(48).reshape(6, 4, 2)
ds = Dataset(data,
sa={
'timepoints': np.arange(6),
'multidim': data.copy()
},
fa={'fid': np.arange(4)})
# with overlapping and non-overlapping boxcars
startpoints = [0, 1, 4]
boxlength = 2
bm = BoxcarMapper(startpoints, boxlength, space='boxy')
# train is critical
bm.train(ds)
mds = bm.forward(ds)
assert_equal(len(mds), len(startpoints))
assert_equal(mds.nfeatures, boxlength)
# all samples attributes remain, but the can rotated/compressed into
# multidimensional attributes
assert_equal(sorted(mds.sa.keys()),
['boxy_onsetidx'] + sorted(ds.sa.keys()))
assert_equal(mds.sa.multidim.shape,
(len(startpoints), boxlength) + ds.shape[1:])
assert_equal(mds.sa.timepoints.shape, (len(startpoints), boxlength))
assert_array_equal(mds.sa.timepoints.flatten(),
np.array([(s, s + 1) for s in startpoints]).flatten())
assert_array_equal(mds.sa.boxy_onsetidx, startpoints)
# feature attributes also get rotated and broadcasted
assert_array_equal(mds.fa.fid, [ds.fa.fid, ds.fa.fid])
# and finally there is a new one
assert_array_equal(mds.fa.boxy_offsetidx, list(range(boxlength)))
# now see how it works on reverse()
rds = bm.reverse(mds)
# we got at least something of all original attributes back
assert_equal(sorted(rds.sa.keys()), sorted(ds.sa.keys()))
assert_equal(sorted(rds.fa.keys()), sorted(ds.fa.keys()))
# it is not possible to reconstruct the full samples array
# some samples even might show up multiple times (when there are overlapping
# boxcars
assert_array_equal(
rds.samples,
np.array([[[0, 1], [2, 3], [4, 5], [6, 7]],
[[8, 9], [10, 11], [12, 13], [14, 15]],
[[8, 9], [10, 11], [12, 13], [14, 15]],
[[16, 17], [18, 19], [20, 21], [22, 23]],
[[32, 33], [34, 35], [36, 37], [38, 39]],
[[40, 41], [42, 43], [44, 45], [46, 47]]]))
assert_array_equal(rds.sa.timepoints, [0, 1, 1, 2, 4, 5])
assert_array_equal(rds.sa.multidim, ds.sa.multidim[rds.sa.timepoints])
# but feature attributes should be fully recovered
assert_array_equal(rds.fa.fid, ds.fa.fid)
# popular dataset configuration (double flatten + boxcar)
cm = ChainMapper([FlattenMapper(), bm, FlattenMapper()])
cm.train(ds)
bflat = ds.get_mapped(cm)
assert_equal(bflat.shape,
(len(startpoints), boxlength * np.prod(ds.shape[1:])))
# add attributes
bflat.fa['testfa'] = np.arange(bflat.nfeatures)
bflat.sa['testsa'] = np.arange(bflat.nsamples)
# now try to go back
bflatrev = bflat.mapper.reverse(bflat)
# data should be same again, as far as the boxcars match
assert_array_equal(ds.samples[:2], bflatrev.samples[:2])
assert_array_equal(ds.samples[-2:], bflatrev.samples[-2:])
# feature axis should match
assert_equal(ds.shape[1:], bflatrev.shape[1:])
def timesegments_classification(dss,
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.
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
...
"""
part2 = NFoldPartitioner(attr='subjects')
# Check if input list contains Datasets, ndarrays
dss = [Dataset(ds) if not type(ds) == Dataset else ds for ds in dss]
# TODO: allow for doing feature selection
if do_zscore:
for ds in dss:
zscore(ds, chunks_attr=None)
# assign .sa.subjects to those datasets
for i, ds in enumerate(dss):
# part2.attr is by default "subjects"
ds.sa[part2.attr] = [i]
dss_test_bc = []
for ds in dss:
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_across_subjects = np.asarray(errors_across_subjects)
if __debug__:
debug(
"BM", "Finished with %s array of errors. Mean error %.2f" %
(errors_across_subjects.shape, np.mean(errors_across_subjects)))
return errors_across_subjects
def test_simpleboxcar():
data = np.atleast_2d(np.arange(10)).T
sp = np.arange(10)
# check if stupid thing don't work
assert_raises(ValueError, BoxcarMapper, sp, 0)
# now do an identity transformation
bcm = BoxcarMapper(sp, 1)
trans = bcm.forward(data)
# ,0 is a feature below, so we get explicit 2D out of 1D
assert_array_equal(trans[:,0], data)
# now check for illegal boxes
if __debug__:
# condition is checked only in __debug__
assert_raises(ValueError, BoxcarMapper(sp, 2).train, data)
# now something that should work
nbox = 9
boxlength = 2
sp = np.arange(nbox)
bcm = BoxcarMapper(sp, boxlength)
trans = bcm.forward(data)
# check that is properly upcasts the dimensionality
assert_equal(trans.shape, (nbox, boxlength) + data.shape[1:])
# check actual values, squeezing the last dim for simplicity
assert_array_equal(trans.squeeze(), np.vstack((np.arange(9), np.arange(9)+1)).T)
# now test for proper data shape
data = np.ones((10,3,4,2))
sp = [ 2, 4, 3, 5 ]
trans = BoxcarMapper(sp, 4).forward(data)
assert_equal(trans.shape, (4,4,3,4,2))
# test reverse
data = np.arange(240).reshape(10, 3, 4, 2)
sp = [ 2, 4, 3, 5 ]
boxlength = 2
m = BoxcarMapper(sp, boxlength)
m.train(data)
mp = m.forward(data)
assert_equal(mp.shape, (4, 2, 3, 4, 2))
# try full reconstruct
mr = m.reverse(mp)
# shape has to match
assert_equal(mr.shape, (len(sp) * boxlength,) + data.shape[1:])
# only known samples are part of the results
assert_true((mr >= 24).all())
assert_true((mr < 168).all())
# check proper reconstruction of non-conflicting sample
assert_array_equal(mr[0].ravel(), np.arange(48, 72))
# check proper reconstruction of samples being part of multiple
# mapped samples
assert_array_equal(mr[1].ravel(), np.arange(72, 96))
# test reverse of a single sample
singlesample = np.arange(48).reshape(2, 3, 4, 2)
assert_array_equal(singlesample, m.reverse1(singlesample))
# now in a dataset
ds = Dataset([singlesample])
assert_equal(ds.shape, (1,) + singlesample.shape)
# after reverse mapping the 'sample axis' should vanish and the original 3d
# shape of the samples should be restored
assert_equal(ds.shape[1:], m.reverse(ds).shape)
# multiple samples should just be concatenated along the samples axis
ds = Dataset([singlesample, singlesample])
assert_equal((np.prod(ds.shape[:2]),) + singlesample.shape[1:],
m.reverse(ds).shape)
# should not work for shape mismatch, but it does work and is useful when
# reverse mapping sample attributes
#assert_raises(ValueError, m.reverse, singlesample[0])
# check broadcasting of 'raw' samples into proper boxcars on forward()
bc = m.forward1(np.arange(24).reshape(3, 4, 2))
assert_array_equal(bc, np.array(2 * [np.arange(24).reshape(3, 4, 2)]))
def test_datasetmapping():
# 6 samples, 4X2 features
data = np.arange(48).reshape(6,4,2)
ds = Dataset(data,
sa={'timepoints': np.arange(6),
'multidim': data.copy()},
fa={'fid': np.arange(4)})
# with overlapping and non-overlapping boxcars
startpoints = [0, 1, 4]
boxlength = 2
bm = BoxcarMapper(startpoints, boxlength, space='boxy')
# train is critical
bm.train(ds)
mds = bm.forward(ds)
assert_equal(len(mds), len(startpoints))
assert_equal(mds.nfeatures, boxlength)
# all samples attributes remain, but the can rotated/compressed into
# multidimensional attributes
assert_equal(sorted(mds.sa.keys()), ['boxy_onsetidx'] + sorted(ds.sa.keys()))
assert_equal(mds.sa.multidim.shape,
(len(startpoints), boxlength) + ds.shape[1:])
assert_equal(mds.sa.timepoints.shape, (len(startpoints), boxlength))
assert_array_equal(mds.sa.timepoints.flatten(),
np.array([(s, s+1) for s in startpoints]).flatten())
assert_array_equal(mds.sa.boxy_onsetidx, startpoints)
# feature attributes also get rotated and broadcasted
assert_array_equal(mds.fa.fid, [ds.fa.fid, ds.fa.fid])
# and finally there is a new one
assert_array_equal(mds.fa.boxy_offsetidx, range(boxlength))
# now see how it works on reverse()
rds = bm.reverse(mds)
# we got at least something of all original attributes back
assert_equal(sorted(rds.sa.keys()), sorted(ds.sa.keys()))
assert_equal(sorted(rds.fa.keys()), sorted(ds.fa.keys()))
# it is not possible to reconstruct the full samples array
# some samples even might show up multiple times (when there are overlapping
# boxcars
assert_array_equal(rds.samples,
np.array([[[ 0, 1], [ 2, 3], [ 4, 5], [ 6, 7]],
[[ 8, 9], [10, 11], [12, 13], [14, 15]],
[[ 8, 9], [10, 11], [12, 13], [14, 15]],
[[16, 17], [18, 19], [20, 21], [22, 23]],
[[32, 33], [34, 35], [36, 37], [38, 39]],
[[40, 41], [42, 43], [44, 45], [46, 47]]]))
assert_array_equal(rds.sa.timepoints, [0, 1, 1, 2, 4, 5])
assert_array_equal(rds.sa.multidim, ds.sa.multidim[rds.sa.timepoints])
# but feature attributes should be fully recovered
assert_array_equal(rds.fa.fid, ds.fa.fid)
# popular dataset configuration (double flatten + boxcar)
cm= ChainMapper([FlattenMapper(), bm, FlattenMapper()])
cm.train(ds)
bflat = ds.get_mapped(cm)
assert_equal(bflat.shape, (len(startpoints), boxlength * np.prod(ds.shape[1:])))
# add attributes
bflat.fa['testfa'] = np.arange(bflat.nfeatures)
bflat.sa['testsa'] = np.arange(bflat.nsamples)
# now try to go back
bflatrev = bflat.mapper.reverse(bflat)
# data should be same again, as far as the boxcars match
assert_array_equal(ds.samples[:2], bflatrev.samples[:2])
assert_array_equal(ds.samples[-2:], bflatrev.samples[-2:])
# feature axis should match
assert_equal(ds.shape[1:], bflatrev.shape[1:])
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 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