def test_hpal_svd_combo(self):
# get seed dataset
ds4l = datasets['uni4large']
ds_orig = ds4l[:, ds4l.a.nonbogus_features]
# XXX Is this SVD mapping required?
svm = SVDMapper()
svm.train(ds_orig)
ds_svs = svm.forward(ds_orig)
ds_orig.samples = ds_svs.samples
nf_true = ds_orig.nfeatures
n = 4 # # of datasets to generate
# Adding non-shared dimensions for each subject
dss_rotated = [[]] * n
for i in range(n):
dss_rotated[i] = hstack(
(ds_orig, ds4l[:, ds4l.a.bogus_features[i * 4:i * 4 + 4]]))
# rotate data
nf = dss_rotated[0].nfeatures
dss_rotated = [
random_affine_transformation(dss_rotated[i]) for i in xrange(n)
]
# Test if it is close to doing hpal+SVD in sequence outside hpal
# First, as we do in sequence outside hpal
ha = Hyperalignment()
mappers_orig = ha(dss_rotated)
dss_back = [
m.forward(ds_) for m, ds_ in zip(mappers_orig, dss_rotated)
]
dss_mean = np.mean([sd.samples for sd in dss_back], axis=0)
svm = SVDMapper()
svm.train(dss_mean)
dss_sv = [svm.forward(sd) for sd in dss_back]
# Test for SVD dimensionality reduction even with 2 training subjects
for output_dim in [1, 4]:
ha = Hyperalignment(output_dim=output_dim)
ha.train(dss_rotated[:2])
mappers = ha(dss_rotated)
dss_back = [m.forward(ds_) for m, ds_ in zip(mappers, dss_rotated)]
for sd in dss_back:
assert (sd.nfeatures == output_dim)
# Check if combined hpal+SVD works as expected
sv_corrs = []
for sd1, sd2 in zip(dss_sv, dss_back):
ndcs = np.diag(np.corrcoef(sd1.samples.T, sd2.samples.T)[nf:, :nf],
k=0)
sv_corrs.append(ndcs)
self.assertTrue(
np.all(np.abs(np.array(sv_corrs)) >= 0.95),
msg="Hyperalignment with dimensionality reduction should have "
"reconstructed SVD dataset. Got correlations %s." % sv_corrs)
# Check if it recovers original SVs
sv_corrs_orig = []
for sd in dss_back:
ndcs = np.diag(np.corrcoef(sd.samples.T,
ds_orig.samples.T)[nf_true:, :nf_true],
k=0)
sv_corrs_orig.append(ndcs)
self.assertTrue(np.all(np.abs(np.array(sv_corrs_orig)) >= 0.9),
msg="Expected original dimensions after "
"SVD. Got correlations %s." % sv_corrs_orig)
def get_testdata(self):
# get a dataset with some prominent trends in it
ds4l = datasets['uni4large']
# lets select for now only meaningful features
ds_orig = ds4l[:, ds4l.a.nonbogus_features]
zscore(ds_orig, chunks_attr=None)
n = 4 # # of datasets to generate
Rs, dss_rotated, dss_rotated_clean = [], [], []
# now lets compose derived datasets by using some random
# rotation(s)
while len(dss_rotated_clean) < n:
ds_ = random_affine_transformation(ds_orig,
scale_fac=1.0,
shift_fac=0.)
if ds_.a.random_scale <= 0:
continue
Rs.append(ds_.a.random_rotation)
zscore(ds_, chunks_attr=None)
dss_rotated_clean.append(ds_)
i = len(dss_rotated_clean) - 1
ds_2 = hstack(
[ds_, ds4l[:, ds4l.a.bogus_features[i * 4:i * 4 + 4]]])
zscore(ds_2, chunks_attr=None)
dss_rotated.append(ds_2)
return ds_orig, dss_rotated, dss_rotated_clean, Rs
def test_random_affine_transformation():
ds = Dataset.from_wizard(np.random.randn(8, 3, 2))
ds_d = random_affine_transformation(ds)
# compare original to the inverse of the distortion using reported
# parameters
assert_array_almost_equal(
np.dot((ds_d.samples - ds_d.a.random_shift) / ds_d.a.random_scale,
ds_d.a.random_rotation.T), ds.samples)
def test_random_affine_transformation():
ds = Dataset.from_wizard(np.random.randn(8,3,2))
ds_d = random_affine_transformation(ds)
# compare original to the inverse of the distortion using reported
# parameters
assert_array_almost_equal(
np.dot((ds_d.samples - ds_d.a.random_shift) / ds_d.a.random_scale,
ds_d.a.random_rotation.T),
ds.samples)
def test_timesegments_classification():
# TODO: RF our construction of fake datasets for testing hyperalignment
# so we could reuse it here and test classification performance
ds_orig = datasets['uni4large']
n = 3
dss = [ds_orig.copy(deep=True) for i in range(n)]
def nohyper(dss):
return [IdentityMapper() for ds in dss]
# clean case, assume "nohyper" which would be by default
errors = timesegments_classification(dss)
for ds in dss:
# must not add any attribute, such as subjects
assert ('subjects' not in ds.sa)
assert_array_equal(errors, 0)
# very noisy case -- we must not be able to classify anything reasonably
dss_noisy = [ds.copy() for ds in dss]
for ds in dss_noisy:
ds.samples = np.random.normal(size=ds.samples.shape)
errors_nonoverlapping = timesegments_classification(
dss_noisy, nohyper, overlapping_windows=False)
assert (np.all(errors_nonoverlapping <= 1.))
assert (np.all(0.75 <= errors_nonoverlapping))
errors_overlapping = timesegments_classification(dss_noisy, nohyper)
# nononverlapping error should be less for random result
assert_array_lequal(np.mean(errors_nonoverlapping),
np.mean(errors_overlapping))
# now the ultimate test with real hyperalignment on when we don't need much
# of it anyways
#import pdb; pdb.set_trace()
dss_rotated = [
random_affine_transformation(ds_orig, scale_fac=100, shift_fac=10)
for _ in dss
]
errors_hyper = timesegments_classification(dss_rotated, Hyperalignment())
# Hyperalignment must not screw up and rotated and classify perfectly
# since we didn't add any noise whatsoever
assert_array_equal(errors, 0)
def test_hpal_joblib(self):
skip_if_no_external('joblib')
# get seed dataset
ds4l = datasets['uni4large']
dss_rotated = [random_affine_transformation(ds4l, scale_fac=100, shift_fac=10)
for i in range(4)]
ha = Hyperalignment(nproc=1, enable_ca=['residual_errors'])
ha.train(dss_rotated[:2])
mappers = ha(dss_rotated)
ha_proc = Hyperalignment(nproc=2, enable_ca=['residual_errors'])
ha_proc.train(dss_rotated[:2])
mappers_nproc = ha_proc(dss_rotated)
# not sure yet why on windows only is not precise
cmp_ = assert_array_equal if (not on_windows) else assert_array_almost_equal
[cmp_(m.proj, mp.proj) for m, mp in zip(mappers, mappers_nproc)] # "Mappers differ when using nproc>1."
cmp_(ha.ca.residual_errors.samples, ha_proc.ca.residual_errors.samples)
# smoke test
ha = Hyperalignment(nproc=0)
mappers = ha(dss_rotated)
def test_hypal_michael_caused_problem(self):
from mvpa2.misc import data_generators
from mvpa2.mappers.zscore import zscore
# Fake data
ds = data_generators.normal_feature_dataset(nfeatures=20)
ds_all = [data_generators.random_affine_transformation(ds) for i in range(3)]
_ = [zscore(sd, chunks_attr=None) for sd in ds_all]
# Making random data per subject for testing with bias added to first subject
ds_test = [np.random.rand(1, ds.nfeatures) for i in range(len(ds_all))]
ds_test[0] += np.arange(1, ds.nfeatures + 1) * 100
assert(np.corrcoef(ds_test[2], ds_test[1])[0, 1] < 0.99) # that would have been rudiculous if it was
# Test with varying alpha so we for sure to not have that issue now
for alpha in (0, 0.01, 0.5, 0.99, 1.0):
hyper09 = Hyperalignment(alpha=alpha)
mappers = hyper09([sd for sd in ds_all])
ds_test_a = [m.forward(sd) for m, sd in zip(mappers, ds_test)]
ds_test_a = [mappers[0].reverse(sd) for sd in ds_test_a]
corr = np.corrcoef(ds_test_a[2], ds_test_a[1])[0, 1]
assert(corr < 0.99)
def test_hpal_joblib(self):
skip_if_no_external('joblib')
# get seed dataset
ds4l = datasets['uni4large']
dss_rotated = [random_affine_transformation(ds4l, scale_fac=100, shift_fac=10)
for i in range(4)]
ha = Hyperalignment(nproc=1, enable_ca=['residual_errors'])
ha.train(dss_rotated[:2])
mappers = ha(dss_rotated)
ha_proc = Hyperalignment(nproc=2, enable_ca=['residual_errors'])
ha_proc.train(dss_rotated[:2])
mappers_nproc = ha_proc(dss_rotated)
self.assertTrue(
np.all([np.array_equal(m.proj, mp.proj)
for m, mp in zip(mappers, mappers_nproc)]),
msg="Mappers differ when using nproc>1.")
assert_array_equal(ha.ca.residual_errors.samples, ha_proc.ca.residual_errors.samples)
# smoke test
ha = Hyperalignment(nproc=0)
mappers = ha(dss_rotated)
def test_hypal_michael_caused_problem(self):
from mvpa2.misc import data_generators
from mvpa2.mappers.zscore import zscore
# Fake data
ds = data_generators.normal_feature_dataset(nfeatures=20)
ds_all = [data_generators.random_affine_transformation(ds) for i in range(3)]
_ = [zscore(sd, chunks_attr=None) for sd in ds_all]
# Making random data per subject for testing with bias added to first subject
ds_test = [np.random.rand(1, ds.nfeatures) for i in range(len(ds_all))]
ds_test[0] += np.arange(1, ds.nfeatures + 1) * 100
assert(np.corrcoef(ds_test[2], ds_test[1])[0, 1] < 0.99) # that would have been ridiculous if it was
# Test with varying alpha so we for sure to not have that issue now
for alpha in (0, 0.01, 0.5, 0.99, 1.0):
hyper09 = Hyperalignment(alpha=alpha)
mappers = hyper09([sd for sd in ds_all])
ds_test_a = [m.forward(sd) for m, sd in zip(mappers, ds_test)]
ds_test_a = [mappers[0].reverse(sd) for sd in ds_test_a]
corr = np.corrcoef(ds_test_a[2], ds_test_a[1])[0, 1]
assert(corr < 0.99)
def test_hpal_joblib(self):
skip_if_no_external('joblib')
# get seed dataset
ds4l = datasets['uni4large']
dss_rotated = [random_affine_transformation(ds4l, scale_fac=100, shift_fac=10)
for i in range(4)]
ha = Hyperalignment(nproc=1, enable_ca=['residual_errors'])
ha.train(dss_rotated[:2])
mappers = ha(dss_rotated)
ha_proc = Hyperalignment(nproc=2, enable_ca=['residual_errors'])
ha_proc.train(dss_rotated[:2])
mappers_nproc = ha_proc(dss_rotated)
self.assertTrue(
np.all([np.array_equal(m.proj, mp.proj)
for m, mp in zip(mappers, mappers_nproc)]),
msg="Mappers differ when using nproc>1.")
assert_array_equal(ha.ca.residual_errors.samples, ha_proc.ca.residual_errors.samples)
# smoke test
ha = Hyperalignment(nproc=0)
mappers = ha(dss_rotated)
def test_timesegments_classification():
# TODO: RF our construction of fake datasets for testing hyperalignment
# so we could reuse it here and test classification performance
ds_orig = datasets['uni4large']
n = 3
dss = [ds_orig.copy(deep=True) for i in xrange(n)]
def nohyper(dss):
return [IdentityMapper() for ds in dss]
# clean case, assume "nohyper" which would be by default
errors = timesegments_classification(dss)
for ds in dss:
# must not add any attribute, such as subjects
assert('subjects' not in ds.sa)
assert_array_equal(errors, 0)
# very noisy case -- we must not be able to classify anything reasonably
dss_noisy = [ds.copy() for ds in dss]
for ds in dss_noisy:
ds.samples = np.random.normal(size=ds.samples.shape)
errors_nonoverlapping = timesegments_classification(dss_noisy, nohyper,
overlapping_windows=False)
assert(np.all(errors_nonoverlapping <= 1.))
assert(np.all(0.85 <= errors_nonoverlapping))
errors_overlapping = timesegments_classification(dss_noisy, nohyper)
# nononverlapping error should be less for random result
assert_array_lequal(np.mean(errors_nonoverlapping), np.mean(errors_overlapping))
# now the ultimate test with real hyperalignment on when we don't need much
# of it anyways
#import pdb; pdb.set_trace()
dss_rotated = [random_affine_transformation(ds_orig, scale_fac=100, shift_fac=10)
for _ in dss]
errors_hyper = timesegments_classification(dss_rotated, Hyperalignment())
# Hyperalignment must not screw up and rotated and classify perfectly
# since we didn't add any noise whatsoever
assert_array_equal(errors, 0)
def get_testdata(self):
# get a dataset with some prominent trends in it
ds4l = datasets['uni4large']
# lets select for now only meaningful features
ds_orig = ds4l[:, ds4l.a.nonbogus_features]
zscore(ds_orig, chunks_attr=None)
n = 4 # # of datasets to generate
Rs, dss_rotated, dss_rotated_clean = [], [], []
# now lets compose derived datasets by using some random
# rotation(s)
while len(dss_rotated_clean) < n:
ds_ = random_affine_transformation(ds_orig, scale_fac=1.0, shift_fac=0.)
if ds_.a.random_scale <= 0:
continue
Rs.append(ds_.a.random_rotation)
zscore(ds_, chunks_attr=None)
dss_rotated_clean.append(ds_)
i = len(dss_rotated_clean) - 1
ds_2 = hstack([ds_, ds4l[:, ds4l.a.bogus_features[i * 4: i * 4 + 4]]])
zscore(ds_2, chunks_attr=None)
dss_rotated.append(ds_2)
return ds_orig, dss_rotated, dss_rotated_clean, Rs
def test_hpal_joblib(self):
skip_if_no_external('joblib')
# get seed dataset
ds4l = datasets['uni4large']
dss_rotated = [
random_affine_transformation(ds4l, scale_fac=100, shift_fac=10)
for i in range(4)
]
ha = Hyperalignment(nproc=1, enable_ca=['residual_errors'])
ha.train(dss_rotated[:2])
mappers = ha(dss_rotated)
ha_proc = Hyperalignment(nproc=2, enable_ca=['residual_errors'])
ha_proc.train(dss_rotated[:2])
mappers_nproc = ha_proc(dss_rotated)
# not sure yet why on windows only is not precise
cmp_ = assert_array_equal if (
not on_windows) else assert_array_almost_equal
[cmp_(m.proj, mp.proj) for m, mp in zip(mappers, mappers_nproc)
] # "Mappers differ when using nproc>1."
cmp_(ha.ca.residual_errors.samples, ha_proc.ca.residual_errors.samples)
# smoke test
ha = Hyperalignment(nproc=0)
mappers = ha(dss_rotated)
def test_basic_functioning(self, ref_ds, zscore_common, zscore_all):
ha = Hyperalignment(ref_ds=ref_ds,
zscore_all=zscore_all,
zscore_common=zscore_common)
if ref_ds is None:
ref_ds = 0 # by default should be this one
# get a dataset with some prominent trends in it
ds4l = datasets['uni4large']
# lets select for now only meaningful features
ds_orig = ds4l[:, ds4l.a.nonbogus_features]
nf = ds_orig.nfeatures
n = 4 # # of datasets to generate
Rs, dss_rotated, dss_rotated_clean, random_shifts, random_scales \
= [], [], [], [], []
# now lets compose derived datasets by using some random
# rotation(s)
for i in xrange(n):
## if False: # i == ref_ds:
# # Do not rotate the target space so we could check later on
# # if we transform back nicely
# R = np.eye(ds_orig.nfeatures)
## else:
ds_ = random_affine_transformation(ds_orig,
scale_fac=100,
shift_fac=10)
Rs.append(ds_.a.random_rotation)
# reusing random data from dataset itself
random_scales += [ds_.a.random_scale]
random_shifts += [ds_.a.random_shift]
random_noise = ds4l.samples[:, ds4l.a.bogus_features[:4]]
## if (zscore_common or zscore_all):
## # for later on testing of "precise" reconstruction
## zscore(ds_, chunks_attr=None)
dss_rotated_clean.append(ds_)
ds_ = ds_.copy()
ds_.samples = ds_.samples + 0.1 * random_noise
dss_rotated.append(ds_)
# Lets test two scenarios -- in one with no noise -- we should get
# close to perfect reconstruction. If noise was added -- not so good
for noisy, dss in ((False, dss_rotated_clean), (True, dss_rotated)):
# to verify that original datasets didn't get changed by
# Hyperalignment store their idhashes of samples
idhashes = [idhash(ds.samples) for ds in dss]
idhashes_targets = [idhash(ds.targets) for ds in dss]
mappers = ha(dss)
idhashes_ = [idhash(ds.samples) for ds in dss]
idhashes_targets_ = [idhash(ds.targets) for ds in dss]
self.assertEqual(
idhashes,
idhashes_,
msg="Hyperalignment must not change original data.")
self.assertEqual(
idhashes_targets,
idhashes_targets_,
msg="Hyperalignment must not change original data targets.")
self.assertEqual(ref_ds, ha.ca.chosen_ref_ds)
# Map data back
dss_clean_back = [
m.forward(ds_) for m, ds_ in zip(mappers, dss_rotated_clean)
]
ds_norm = np.linalg.norm(dss[ref_ds].samples)
nddss = []
ndcss = []
ds_orig_Rref = np.dot(ds_orig.samples, Rs[ref_ds]) \
* random_scales[ref_ds] \
+ random_shifts[ref_ds]
if zscore_common or zscore_all:
zscore(Dataset(ds_orig_Rref), chunks_attr=None)
for ds_back in dss_clean_back:
# if we used zscoring of common, we cannot rely
# that range/offset could be matched, so lets use
# corrcoef
ndcs = np.diag(np.corrcoef(ds_back.samples.T,
ds_orig_Rref.T)[nf:, :nf],
k=0)
ndcss += [ndcs]
dds = ds_back.samples - ds_orig_Rref
ndds = np.linalg.norm(dds) / ds_norm
nddss += [ndds]
snoisy = ('clean', 'noisy')[int(noisy)]
do_labile = cfg.getboolean('tests', 'labile', default='yes')
if not noisy or do_labile:
# First compare correlations
self.assertTrue(
np.all(np.array(ndcss) >= (0.9, 0.85)[int(noisy)]),
msg="Should have reconstructed original dataset more or"
" less. Got correlations %s in %s case." % (ndcss, snoisy))
if not (zscore_all or zscore_common):
# if we didn't zscore -- all of them should be really close
self.assertTrue(
np.all(np.array(nddss) <= (1e-10, 1e-1)[int(noisy)]),
msg="Should have reconstructed original dataset well "
"without zscoring. Got normed differences %s in %s case."
% (nddss, snoisy))
elif do_labile:
# otherwise they all should be somewhat close
self.assertTrue(
np.all(np.array(nddss) <= (.2, 3)[int(noisy)]),
msg="Should have reconstructed original dataset more or"
" less for all. Got normed differences %s in %s case."
% (nddss, snoisy))
self.assertTrue(
np.all(nddss[ref_ds] <= .09),
msg="Should have reconstructed original dataset quite "
"well even with zscoring. Got normed differences %s "
"in %s case." % (nddss, snoisy))
# yoh: and leave 5% of difference for a chance and numerical
# fluctuations ;)
self.assertTrue(
np.all(np.array(nddss) >= 0.95 * nddss[ref_ds]),
msg="Should have reconstructed orig_ds best of all. "
"Got normed differences %s in %s case with ref_ds=%d."
% (nddss, snoisy, ref_ds))
# Lets see how well we do if asked to compute residuals
ha = Hyperalignment(
ref_ds=ref_ds,
level2_niter=2,
enable_ca=['training_residual_errors', 'residual_errors'])
mappers = ha(dss_rotated_clean)
self.assertTrue(
np.all(ha.ca.training_residual_errors.sa.levels ==
['1', '2:0', '2:1']))
rterrors = ha.ca.training_residual_errors.samples
# just basic tests:
self.assertEqual(rterrors[0, ref_ds], 0)
self.assertEqual(rterrors.shape, (3, n))
rerrors = ha.ca.residual_errors.samples
self.assertEqual(rerrors.shape, (1, n))
def test_basic_functioning(self, ref_ds, zscore_common, zscore_all):
ha = Hyperalignment(ref_ds=ref_ds,
zscore_all=zscore_all,
zscore_common=zscore_common)
if ref_ds is None:
ref_ds = 0 # by default should be this one
# get a dataset with some prominent trends in it
ds4l = datasets['uni4large']
# lets select for now only meaningful features
ds_orig = ds4l[:, ds4l.a.nonbogus_features]
nf = ds_orig.nfeatures
n = 4 # # of datasets to generate
Rs, dss_rotated, dss_rotated_clean, random_shifts, random_scales \
= [], [], [], [], []
# now lets compose derived datasets by using some random
# rotation(s)
for i in xrange(n):
## if False: # i == ref_ds:
# # Do not rotate the target space so we could check later on
# # if we transform back nicely
# R = np.eye(ds_orig.nfeatures)
## else:
ds_ = random_affine_transformation(ds_orig, scale_fac=100, shift_fac=10)
Rs.append(ds_.a.random_rotation)
# reusing random data from dataset itself
random_scales += [ds_.a.random_scale]
random_shifts += [ds_.a.random_shift]
random_noise = ds4l.samples[:, ds4l.a.bogus_features[:4]]
## if (zscore_common or zscore_all):
## # for later on testing of "precise" reconstruction
## zscore(ds_, chunks_attr=None)
dss_rotated_clean.append(ds_)
ds_ = ds_.copy()
ds_.samples = ds_.samples + 0.1 * random_noise
dss_rotated.append(ds_)
# Lets test two scenarios -- in one with no noise -- we should get
# close to perfect reconstruction. If noise was added -- not so good
for noisy, dss in ((False, dss_rotated_clean),
(True, dss_rotated)):
# to verify that original datasets didn't get changed by
# Hyperalignment store their idhashes of samples
idhashes = [idhash(ds.samples) for ds in dss]
idhashes_targets = [idhash(ds.targets) for ds in dss]
mappers = ha(dss)
idhashes_ = [idhash(ds.samples) for ds in dss]
idhashes_targets_ = [idhash(ds.targets) for ds in dss]
self.assertEqual(idhashes, idhashes_,
msg="Hyperalignment must not change original data.")
self.assertEqual(idhashes_targets, idhashes_targets_,
msg="Hyperalignment must not change original data targets.")
self.assertEqual(ref_ds, ha.ca.chosen_ref_ds)
# Map data back
dss_clean_back = [m.forward(ds_)
for m, ds_ in zip(mappers, dss_rotated_clean)]
ds_norm = np.linalg.norm(dss[ref_ds].samples)
nddss = []
ndcss = []
ds_orig_Rref = np.dot(ds_orig.samples, Rs[ref_ds]) \
* random_scales[ref_ds] \
+ random_shifts[ref_ds]
if zscore_common or zscore_all:
zscore(Dataset(ds_orig_Rref), chunks_attr=None)
for ds_back in dss_clean_back:
# if we used zscoring of common, we cannot rely
# that range/offset could be matched, so lets use
# corrcoef
ndcs = np.diag(np.corrcoef(ds_back.samples.T,
ds_orig_Rref.T)[nf:, :nf], k=0)
ndcss += [ndcs]
dds = ds_back.samples - ds_orig_Rref
ndds = np.linalg.norm(dds) / ds_norm
nddss += [ndds]
snoisy = ('clean', 'noisy')[int(noisy)]
do_labile = cfg.getboolean('tests', 'labile', default='yes')
if not noisy or do_labile:
# First compare correlations
self.assertTrue(np.all(np.array(ndcss)
>= (0.9, 0.85)[int(noisy)]),
msg="Should have reconstructed original dataset more or"
" less. Got correlations %s in %s case."
% (ndcss, snoisy))
if not (zscore_all or zscore_common):
# if we didn't zscore -- all of them should be really close
self.assertTrue(np.all(np.array(nddss)
<= (1e-10, 1e-1)[int(noisy)]),
msg="Should have reconstructed original dataset well "
"without zscoring. Got normed differences %s in %s case."
% (nddss, snoisy))
elif do_labile:
# otherwise they all should be somewhat close
self.assertTrue(np.all(np.array(nddss)
<= (.2, 3)[int(noisy)]),
msg="Should have reconstructed original dataset more or"
" less for all. Got normed differences %s in %s case."
% (nddss, snoisy))
self.assertTrue(np.all(nddss[ref_ds] <= .09),
msg="Should have reconstructed original dataset quite "
"well even with zscoring. Got normed differences %s "
"in %s case." % (nddss, snoisy))
# yoh: and leave 5% of difference for a chance and numerical
# fluctuations ;)
self.assertTrue(np.all(np.array(nddss) >= 0.95*nddss[ref_ds]),
msg="Should have reconstructed orig_ds best of all. "
"Got normed differences %s in %s case with ref_ds=%d."
% (nddss, snoisy, ref_ds))
# Lets see how well we do if asked to compute residuals
ha = Hyperalignment(ref_ds=ref_ds, level2_niter=2,
enable_ca=['training_residual_errors',
'residual_errors'])
mappers = ha(dss_rotated_clean)
self.assertTrue(np.all(ha.ca.training_residual_errors.sa.levels ==
['1', '2:0', '2:1']))
rterrors = ha.ca.training_residual_errors.samples
# just basic tests:
self.assertEqual(rterrors[0, ref_ds], 0)
self.assertEqual(rterrors.shape, (3, n))
rerrors = ha.ca.residual_errors.samples
self.assertEqual(rerrors.shape, (1, n))
def test_hpal_svd_combo(self):
# get seed dataset
ds4l = datasets['uni4large']
ds_orig = ds4l[:, ds4l.a.nonbogus_features]
# XXX Is this SVD mapping required?
svm = SVDMapper()
svm.train(ds_orig)
ds_svs = svm.forward(ds_orig)
ds_orig.samples = ds_svs.samples
nf_true = ds_orig.nfeatures
n = 4 # # of datasets to generate
# Adding non-shared dimensions for each subject
dss_rotated = [[]]*n
for i in range(n):
dss_rotated[i] = hstack(
(ds_orig, ds4l[:, ds4l.a.bogus_features[i * 4: i * 4 + 4]]))
# rotate data
nf = dss_rotated[0].nfeatures
dss_rotated = [random_affine_transformation(dss_rotated[i])
for i in xrange(n)]
# Test if it is close to doing hpal+SVD in sequence outside hpal
# First, as we do in sequence outside hpal
ha = Hyperalignment()
mappers_orig = ha(dss_rotated)
dss_back = [m.forward(ds_)
for m, ds_ in zip(mappers_orig, dss_rotated)]
dss_mean = np.mean([sd.samples for sd in dss_back], axis=0)
svm = SVDMapper()
svm.train(dss_mean)
dss_sv = [svm.forward(sd) for sd in dss_back]
# Test for SVD dimensionality reduction even with 2 training subjects
for output_dim in [1, 4]:
ha = Hyperalignment(output_dim=output_dim)
ha.train(dss_rotated[:2])
mappers = ha(dss_rotated)
dss_back = [m.forward(ds_)
for m, ds_ in zip(mappers, dss_rotated)]
for sd in dss_back:
assert (sd.nfeatures == output_dim)
# Check if combined hpal+SVD works as expected
sv_corrs = []
for sd1, sd2 in zip(dss_sv, dss_back):
ndcs = np.diag(np.corrcoef(sd1.samples.T, sd2.samples.T)[nf:, :nf],
k=0)
sv_corrs.append(ndcs)
self.assertTrue(
np.all(np.abs(np.array(sv_corrs)) >= 0.95),
msg="Hyperalignment with dimensionality reduction should have "
"reconstructed SVD dataset. Got correlations %s."
% sv_corrs)
# Check if it recovers original SVs
sv_corrs_orig = []
for sd in dss_back:
ndcs = np.diag(
np.corrcoef(sd.samples.T, ds_orig.samples.T)[nf_true:, :nf_true],
k=0)
sv_corrs_orig.append(ndcs)
self.assertTrue(
np.all(np.abs(np.array(sv_corrs_orig)) >= 0.9),
msg="Expected original dimensions after "
"SVD. Got correlations %s."
% sv_corrs_orig)