def test_pgso():
Xs = next(generate_mcca_test_data())
N = Xs[0].shape[0]
ranks = []
corrs = []
for tol in [0, 0.1, 0.5, 1]:
kmcca = KMCCA(kernel='rbf', pgso=True, tol=tol, n_components=2)
scores = kmcca.fit(Xs).transform(Xs)
# kmcca2 = KMCCA(kernel='rbf', pgso=True, tol=tol)
# scores2 = kmcca2.fit(Xs).transform(Xs)
for v in range(len(Xs)):
assert len(
set([
kmcca.pgso_Ls_[v].shape[1], kmcca.pgso_norms_[v].shape[0],
kmcca.pgso_idxs_[v].shape[0], kmcca.pgso_Xs_[v].shape[0]
])) == 1
R = kmcca._get_kernel(Xs[v], v) - \
kmcca.pgso_Ls_[v] @ kmcca.pgso_Ls_[v].T
R = R - kmcca.kernel_col_means_[v] - \
kmcca.kernel_col_means_[v].T + kmcca.kernel_mat_means_[v]
assert np.trace(R) / N <= tol
ranks.append(kmcca.pgso_ranks_)
corrs.append(kmcca.canon_corrs(scores))
if tol == 0:
assert np.all(ranks[-1] == [N for _ in Xs])
assert np.all(np.diff(corrs, axis=0) <= 1e-10), corrs
assert np.all(np.diff(corrs, axis=1) <= 1e-10), corrs
assert np.all(np.diff(ranks, axis=0) <= 1e-10), ranks
def test_kmcca_n_components():
Xs = next(generate_mcca_test_data())
n_features = sum([X.shape[1] for X in Xs])
kmcca = KMCCA(n_components=n_features + 1)
with pytest.raises(AttributeError):
kmcca.n_components_
with pytest.warns(None):
kmcca.fit(Xs)
assert kmcca.n_components_ == n_features
for load in kmcca.dual_vars_:
assert kmcca.n_components_ == load.shape[1]
def test_kmcca_params(kernel_args, diag_mode, signal_ranks, n_components,
regs, sval_thresh, center, multiview_output):
kernel, kernel_params = kernel_args
kmcca = KMCCA(
kernel=kernel, kernel_params=kernel_params,
diag_mode=diag_mode, signal_ranks=signal_ranks,
n_components=n_components, regs=regs, sval_thresh=sval_thresh,
center=center, multiview_output=multiview_output)
Xs = next(generate_mcca_test_data())
scores = kmcca.fit_transform(Xs)
if multiview_output:
assert len(scores) == len(Xs)
canon_corrs = kmcca.canon_corrs(scores)
assert len(canon_corrs) == scores.shape[2]
else:
assert scores.shape == (Xs[0].shape[0], kmcca.n_components_)
def test_kmcca(diag_mode):
for Xs in generate_mcca_test_data():
for params in generate_mcca_test_settings():
if len(Xs) == 2 and params['n_components'] is None:
# this setting raises some issues where are few
# of the view scores are not equal. I do not think
# this is an issue in practice so lets just skip
# this scenario
continue
# Linear KMCCA should match MCCA
n_features = [x.shape[1] for x in Xs]
kmcca = KMCCA(kernel='linear',
sval_thresh=0,
signal_ranks=n_features,
diag_mode=diag_mode,
**params).fit(Xs)
mcca = MCCA(**params).fit(Xs)
# evals
np.testing.assert_array_almost_equal(kmcca.evals_,
mcca.evals_,
decimal=2)
# MCCA and linear KMCCA output should be the same
for ks, ms in zip(kmcca.transform(Xs), mcca.transform(Xs)):
np.testing.assert_array_almost_equal(ks, ms, decimal=2)
# Linear KMCCA should match MCCA
n_features = [x.shape[1] for x in Xs]
kmcca = KMCCA(kernel='linear',
sval_thresh=0,
signal_ranks=n_features,
diag_mode=diag_mode,
multiview_output=False,
**params).fit(Xs)
mcca = MCCA(multiview_output=False, **params).fit(Xs)
# evals
np.testing.assert_array_almost_equal(kmcca.evals_,
mcca.evals_,
decimal=2)
# common_scores_normed
np.testing.assert_array_almost_equal(kmcca.transform(Xs),
mcca.transform(Xs),
decimal=2)