Skip to content

[MRG] FIX: fix lsml inversion #203

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 3 commits into from
Closed

[MRG] FIX: fix lsml inversion #203

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
b958e19
FIX: fix lsml inversion
May 13, 2019
89f22b2
Add todo
May 13, 2019
90a4c04
Add better test
May 13, 2019
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 4 additions & 4 deletions metric_learn/lsml.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -39,7 +39,7 @@ def __init__(self, tol=1e-3, max_iter=1000, prior=None, verbose=False,
tol : float, optional
max_iter : int, optional
prior : (d x d) matrix, optional
guess at a metric [default: inv(covariance(X))]
guess at a metric [default: pinvh(covariance(X))]
verbose : bool, optional
if True, prints information while learning
preprocessor : array-like, shape=(n_samples, n_features) or callable
Expand Down Expand Up @@ -70,10 +70,10 @@ def _fit(self, quadruplets, weights=None):
X = np.vstack({tuple(row) for row in
quadruplets.reshape(-1, quadruplets.shape[2])})
prior_inv = np.atleast_2d(np.cov(X, rowvar=False))
M = np.linalg.inv(prior_inv)
M = scipy.linalg.pinvh(prior_inv)
else:
M = self.prior
prior_inv = np.linalg.inv(self.prior)
prior_inv = scipy.linalg.pinvh(self.prior)

step_sizes = np.logspace(-10, 0, 10)
# Keep track of the best step size and the loss at that step.
Expand Down Expand Up @@ -126,7 +126,7 @@ def _total_loss(self, metric, vab, vcd, prior_inv):
return self._comparison_loss(metric, vab, vcd) + reg_loss

def _gradient(self, metric, vab, vcd, prior_inv):
dMetric = prior_inv - np.linalg.inv(metric)
dMetric = prior_inv - scipy.linalg.pinvh(metric)
dabs = np.sum(vab.dot(metric) * vab, axis=1)
dcds = np.sum(vcd.dot(metric) * vcd, axis=1)
violations = dabs > dcds
Expand Down
12 changes: 12 additions & 0 deletions test/metric_learn_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -74,6 +74,18 @@ def test_deprecation_num_labeled(self):
'removed in 0.6.0')
assert_warns_message(DeprecationWarning, msg, lsml_supervised.fit, X, y)

def test_singular_covariance_does_not_diverge(self):
# Test that LSML does not diverge when using the covariance prior and
# when this covariance has null eigenvalues (See
# https://github.com/metric-learn/metric-learn/issues/202)
# TODO: remove when # 195 is merged
rng = np.random.RandomState(42)
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

somehow the covariance matrix in the previous example was not exactly singular so I built the example by hand, and it is indeed truly singular and exhibited another place where we should use the pseudo inverse

X, y = load_iris(return_X_y=True)
# we add a feature that is a linear combination of the two first ones
X = np.concatenate([X, X[:, :2].dot(rng.randn(2, 1))], axis=-1)
lsml = LSML_Supervised()
lsml.fit(X, y, random_state=rng)


class TestITML(MetricTestCase):
def test_iris(self):
Expand Down