Refactor tests (#136)

* rewriting base file to get ride of numpy matrices. * last modification to rewrite into full numpy integration * rewriting mean with dotted notation and getting rid of numpy matrices * continuing to rewrite the geodesic * rewriting of the spatial filters * small style correction in channelselection * updating the use of numpy for all higher packages * last rewriting of pyriemann file * adding slack notification * finishing to update numpy ref and removing numpy matrices * covering all functions of classification: TSclassifier, kNN * PEP8 conformation for test clustering * PEP 8 normalisation * fixing last bugs for testing * update test * adding test for Potato * end of modification of test units and coverage improvment. * modifying and adapting travis file. * updating travis file and correcting error in testing wasserstein mean * adding pandas in the requierement for Travis * ajout de la generation de matrice SPD * Correcting the Karcher mean implementation if the gradient descent is initialized on one of the point of the set, generating zero-divide error. * adding ALM to mean computation * adding ref for the parametrized geodesic * update avant merge * uncomplete merge * merge with upstream * use fixture and parametrize for tests, avoid code redundancy, systematic test of dist and metric * correct tangent fit bug Co-authored-by: kjersbry <noreply@github.com> Co-authored-by: sylvchev <noreply@github.com> * add tsupdate check * reduce complexity of the test to make them faster * add flake8 to tests and examples, correct style in tests * switch to triangular inequality to avoid random failure * update coverage * remove redundant tests, add coverage * correct flake8 * correct F401,F811 import errors with fixtures * correct w value * remove assert_array_equal and correct variable names Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * add is_spsd and get_labels fixture * adding test for Schaefer estimator * update test for spatial filters Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * correct var name for inverse transform * use get_labels fixture * correct flake8 * add is_psd and is_semi_psd * add independence test for fit/transform * correct zeros -> empty * Schaefer-Strimmer shrinkage covariance estimator (#59) * rewriting base file to get ride of numpy matrices. * last modification to rewrite into full numpy integration * rewriting mean with dotted notation and getting rid of numpy matrices * continuing to rewrite the geodesic * rewriting of the spatial filters * small style correction in channelselection * updating the use of numpy for all higher packages * last rewriting of pyriemann file * adding slack notification * finishing to update numpy ref and removing numpy matrices * covering all functions of classification: TSclassifier, kNN * PEP8 conformation for test clustering * PEP 8 normalisation * fixing last bugs for testing * update test * adding test for Potato * end of modification of test units and coverage improvment. * modifying and adapting travis file. * updating travis file and correcting error in testing wasserstein mean * adding pandas in the requierement for Travis * ajout de la generation de matrice SPD * Correcting the Karcher mean implementation if the gradient descent is initialized on one of the point of the set, generating zero-divide error. * adding ALM to mean computation * adding ref for the parametrized geodesic * update avant merge * adding computation and test of Schaefer-Strimmer covariance estimator * replacing loops with algebra * adding an example * correcting plots * uncomplete merge * pep8 * update test for tangent space * correct docstring and bug * converting notebook into py file * correct style * rewrite plots, reduce computation time * update link and rename variables * correct style * correct plot layout * adding Schaefer-Strimmer in whatsnew * update test, docstring and docs from suggestions Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> Co-authored-by: Alexandre Gramfort <alexandre.gramfort@m4x.org> * improve tests for covariance * keep the same color for each estimator across figures * correct example text Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * raise error for covariance_EP and initialize RandomState in example * correct import order * Apply suggestions from code review Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * sort import, correct spd Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * switch zeros to np.empty Co-authored-by: Alexandre Gramfort <alexandre.gramfort@m4x.org> Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> Co-authored-by: Alexandre Gramfort <alexandre.gramfort@m4x.org> * add doc * rebase on master * correct tangent fit bug Co-authored-by: kjersbry <noreply@github.com> Co-authored-by: sylvchev <noreply@github.com> * add tsupdate check * reduce complexity of the test to make them faster * add flake8 to tests and examples, correct style in tests * switch to triangular inequality to avoid random failure * update coverage * remove redundant tests, add coverage * correct flake8 * correct F401,F811 import errors with fixtures * correct w value * add is_spsd and get_labels fixture * remove assert_array_equal and correct variable names Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * adding test for Schaefer estimator * correct var name for inverse transform * update test for spatial filters Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> * use get_labels fixture * correct flake8 * add is_psd and is_semi_psd * add independence test for fit/transform * correct zeros -> empty * define is_spd/is_spsd fixtures * clean code * updating whats new Co-authored-by: kjersbry <noreply@github.com> Co-authored-by: Quentin Barthélemy <q.barthelemy@gmail.com> Co-authored-by: Alexandre Gramfort <alexandre.gramfort@m4x.org>
pyRiemann · Sep 20, 2021 · 9e12432 · 9e12432
1 parent 3ac67b7
commit 9e12432
Show file tree

Hide file tree

Showing 22 changed files with 1,527 additions and 1,094 deletions.
diff --git a/.github/workflows/testing.yml b/.github/workflows/testing.yml
@@ -33,7 +33,7 @@ jobs:
         python -m pip install .[tests]
     - name: Lint with flake8
       run: |
-        flake8 examples pyriemann
+        flake8 examples tests pyriemann
     - name: Test with pytest
       run: |
         pytest
diff --git a/doc/whatsnew.rst b/doc/whatsnew.rst
@@ -22,6 +22,8 @@ v0.2.8.dev
 
 - Add Schaefer-Strimmer covariance estimator in :func:`pyriemann.utils.covariance.covariances`, and an example to compare estimators
 
+- Refactor tests + fix refit of :class:`pyriemann.tangentspace.TangentSpace`
+
 v0.2.7 (June 2021)
 ------------------
 

diff --git a/pyriemann/clustering.py b/pyriemann/clustering.py
@@ -80,15 +80,12 @@ class Kmeans(BaseEstimator, ClassifierMixin, ClusterMixin, TransformerMixin):
         The generator used to initialize the centers. If an integer is
         given, it fixes the seed. Defaults to the global numpy random
         number generator.
-    init : 'k-means++', 'random' or an ndarray (default 'random')
+    init : 'random' or an ndarray (default 'random')
         Method for initialization of centers.
-        'k-means++' : selects initial cluster centers for k-mean
-        clustering in a smart way to speed up convergence. See section
-        Notes in k_init for more details.
         'random': choose k observations (rows) at random from data for
         the initial centroids.
-        If an ndarray is passed, it should be of shape (n_clusters, n_features)
-        and gives the initial centers.
+        If an ndarray is passed, it should be of shape
+        (n_clusters, n_channels, n_channels) and gives the initial centers.
     n_init : int, (default: 10)
         Number of time the k-means algorithm will be run with different
         centroid seeds. The final results will be the best output of
@@ -151,7 +148,7 @@ def fit(self, X, y=None):
         self : Kmeans instance
             The Kmean instance.
         """
-        if (self.init != 'random') | (self.n_init == 1):
+        if (self.init != 'random'):
             # no need to iterate if init is not random
             labels, inertia, mdm = _fit_single(X, y,
                                                n_clusters=self.n_clusters,

diff --git a/pyriemann/spatialfilters.py b/pyriemann/spatialfilters.py
@@ -185,7 +185,7 @@ def __init__(self, filters, log=False):
         self.log = log
 
     def fit(self, X, y):
-        """Train CSP spatial filters.
+        """Train BilinearFilter spatial filters.
 
         Parameters
         ----------
@@ -196,8 +196,8 @@ def fit(self, X, y):
 
         Returns
         -------
-        self : CSP instance
-            The CSP instance.
+        self : BilinearFilter instance
+            The BilinearFilter instance.
         """
         self.filters_ = self.filters
         return self

diff --git a/pyriemann/tangentspace.py b/pyriemann/tangentspace.py
@@ -164,7 +164,6 @@ def fit_transform(self, X, y=None, sample_weight=None):
             the tangent space projection of the matrices.
         """
         # compute mean covariance
-        self._check_reference_points(X)
         self.reference_ = mean_covariance(X, metric=self.metric,
                                           sample_weight=sample_weight)
         return tangent_space(X, self.reference_)

diff --git a/tests/conftest.py b/tests/conftest.py
@@ -1,4 +1,5 @@
 import pytest
+from pytest import approx
 import numpy as np
 from functools import partial
 
@@ -22,27 +23,178 @@ def requires_module(function, name, call=None):
 requires_seaborn = partial(requires_module, name="seaborn")
 
 
-def generate_cov(n_trials, n_channels):
-    """Generate a set of cavariances matrices for test purpose"""
-    rs = np.random.RandomState(1234)
+def generate_cov(n_trials, n_channels, rs, return_params=False):
+    """Generate a set of covariances matrices for test purpose"""
     diags = 2.0 + 0.1 * rs.randn(n_trials, n_channels)
     A = 2 * rs.rand(n_channels, n_channels) - 1
     A /= np.linalg.norm(A, axis=1)[:, np.newaxis]
     covmats = np.empty((n_trials, n_channels, n_channels))
     for i in range(n_trials):
         covmats[i] = A @ np.diag(diags[i]) @ A.T
-    return covmats, diags, A
+    if return_params:
+        return covmats, diags, A
+    else:
+        return covmats
+
+
+@pytest.fixture
+def rndstate():
+    return np.random.RandomState(1234)
+
+
+@pytest.fixture
+def get_covmats(rndstate):
+    def _gen_cov(n_trials, n_chan):
+        return generate_cov(n_trials, n_chan, rndstate, return_params=False)
+
+    return _gen_cov
+
+
+@pytest.fixture
+def get_covmats_params(rndstate):
+    def _gen_cov_params(n_trials, n_chan):
+        return generate_cov(n_trials, n_chan, rndstate, return_params=True)
+
+    return _gen_cov_params
+
+
+@pytest.fixture
+def get_labels():
+    def _get_labels(n_trials, n_classes):
+        return np.arange(n_classes).repeat(n_trials // n_classes)
+
+    return _get_labels
+
+
+def is_positive_semi_definite(X):
+    """Check if all matrices are positive semi-definite.
+
+    Parameters
+    ----------
+    X : ndarray, shape (..., n, n)
+        The set of square matrices, at least 2D ndarray.
+
+    Returns
+    -------
+    ret : boolean
+        True if all matrices are positive semi-definite.
+    """
+    cs = X.shape[-1]
+    return np.all(np.linalg.eigvals(X.reshape((-1, cs, cs))) >= 0.0)
+
+
+def is_positive_definite(X):
+    """Check if all matrices are positive definite.
+
+    Parameters
+    ----------
+    X : ndarray, shape (..., n, n)
+        The set of square matrices, at least 2D ndarray.
+
+    Returns
+    -------
+    ret : boolean
+        True if all matrices are positive definite.
+    """
+    cs = X.shape[-1]
+    return np.all(np.linalg.eigvals(X.reshape((-1, cs, cs))) > 0.0)
+
+
+def is_symmetric(X):
+    """Check if all matrices are symmetric.
+
+    Parameters
+    ----------
+    X : ndarray, shape (..., n, n)
+        The set of square matrices, at least 2D ndarray.
+
+    Returns
+    -------
+    ret : boolean
+        True if all matrices are symmetric.
+    """
+    return X == approx(np.swapaxes(X, -2, -1))
 
 
 @pytest.fixture
-def covmats():
-    """Generate covariance matrices for test"""
-    covmats, _, _ = generate_cov(6, 3)
-    return covmats
+def is_spd():
+    """Check if all matrices are symmetric positive-definite.
+
+    Parameters
+    ----------
+    X : ndarray, shape (..., n, n)
+        The set of square matrices, at least 2D ndarray.
+
+    Returns
+    -------
+    ret : boolean
+        True if all matrices are symmetric positive-definite.
+    """
+
+    def _is_spd(X):
+        return is_symmetric(X) and is_positive_definite(X)
+
+    return _is_spd
 
 
 @pytest.fixture
-def many_covmats():
-    """Generate covariance matrices for test"""
-    covmats, _, _ = generate_cov(100, 3)
-    return covmats
+def is_spsd():
+    """Check if all matrices are symmetric positive semi-definite.
+
+    Parameters
+    ----------
+    X : ndarray, shape (..., n, n)
+        The set of square matrices, at least 2D ndarray.
+
+    Returns
+    -------
+    ret : boolean
+        True if all matrices are symmetric positive semi-definite.
+    """
+
+    def _is_spsd(X):
+        return is_symmetric(X) and is_positive_semi_definite(X)
+
+    return _is_spsd
+
+
+def get_distances():
+    distances = [
+        "riemann",
+        "logeuclid",
+        "euclid",
+        "logdet",
+        "kullback",
+        "kullback_right",
+        "kullback_sym",
+    ]
+    for dist in distances:
+        yield dist
+
+
+def get_means():
+    means = [
+        "riemann",
+        "logeuclid",
+        "euclid",
+        "logdet",
+        "identity",
+        "wasserstein",
+        "ale",
+        "harmonic",
+        "kullback_sym",
+    ]
+    for mean in means:
+        yield mean
+
+
+def get_metrics():
+    metrics = [
+        "riemann",
+        "logeuclid",
+        "euclid",
+        "logdet",
+        "kullback_sym",
+    ]
+    for met in metrics:
+        yield met
diff --git a/tests/test_ajd.py b/tests/test_ajd.py
@@ -1,73 +1,90 @@
 from numpy.testing import assert_array_equal
 import numpy as np
-from numpy.testing import assert_allclose
 import pytest
+from pytest import approx
 
 from pyriemann.utils.ajd import rjd, ajd_pham, uwedge, _get_normalized_weight
 
 
-def generate_cov(Nt, Ne):
-    """Generate a set of cavariances matrices for test purpose"""
-    rs = np.random.RandomState(1234)
-    diags = 2.0 + 0.1 * rs.randn(Nt, Ne)
-    A = 2*rs.rand(Ne, Ne) - 1
-    A /= np.atleast_2d(np.sqrt(np.sum(A**2, 1))).T
-    covmats = np.empty((Nt, Ne, Ne))
-    for i in range(Nt):
-        covmats[i] = np.dot(np.dot(A, np.diag(diags[i])), A.T)
-    return covmats, diags, A
-
-
-def test_get_normalized_weight():
+def test_get_normalized_weight(get_covmats):
     """Test get_normalized_weight"""
-    Nt = 100
-    covmats, diags, A = generate_cov(Nt, 3)
+    n_trials, n_channels = 5, 3
+    covmats = get_covmats(n_trials, n_channels)
     w = _get_normalized_weight(None, covmats)
-    assert np.isclose(np.sum(w), 1., atol=1e-10)
+    assert np.sum(w) == approx(1.0, abs=1e-10)
+
 
+def test_get_normalized_weight_length(get_covmats):
+    n_trials, n_channels = 5, 3
+    covmats = get_covmats(n_trials, n_channels)
+    w = _get_normalized_weight(None, covmats)
     with pytest.raises(ValueError):  # not same length
-        _get_normalized_weight(w[:Nt//2], covmats)
+        _get_normalized_weight(w[: n_trials // 2], covmats)
+
+
+def test_get_normalized_weight_pos(get_covmats):
+    n_trials, n_channels = 5, 3
+    covmats = get_covmats(n_trials, n_channels)
+    w = _get_normalized_weight(None, covmats)
     with pytest.raises(ValueError):  # not strictly positive weight
         w[0] = 0
         _get_normalized_weight(w, covmats)
 
 
-def test_rjd():
-    """Test rjd"""
-    covmats, _, _ = generate_cov(100, 3)
+@pytest.mark.parametrize("ajd", [rjd, ajd_pham])
+def test_ajd_shape(ajd, get_covmats):
+    n_trials, n_channels = 5, 3
+    covmats = get_covmats(n_trials, n_channels)
     V, D = rjd(covmats)
-    assert V.shape == (3, 3)
-    assert D.shape == (100, 3, 3)
+    assert V.shape == (n_channels, n_channels)
+    assert D.shape == (n_trials, n_channels, n_channels)
 
 
-def test_pham():
+def test_pham(get_covmats):
     """Test pham's ajd"""
-    Nt = 100
-    covmats, diags, A = generate_cov(Nt, 3)
+    n_trials, n_channels, w_val = 5, 3, 2
+    covmats = get_covmats(n_trials, n_channels)
     V, D = ajd_pham(covmats)
+    assert V.shape == (n_channels, n_channels)
+    assert D.shape == (n_trials, n_channels, n_channels)
 
-    w = 5 * np.ones(Nt)
-    Vw, Dw = ajd_pham(covmats, sample_weight=w)
+    Vw, Dw = ajd_pham(covmats, sample_weight=w_val * np.ones(n_trials))
     assert_array_equal(V, Vw)  # same result as ajd_pham without weight
     assert_array_equal(D, Dw)
 
+
+def test_pham_pos_weight(get_covmats):
     # Test that weight must be strictly positive
+    n_trials, n_channels, w_val = 5, 3, 2
+    covmats = get_covmats(n_trials, n_channels)
+    w = w_val * np.ones(n_trials)
     with pytest.raises(ValueError):  # not strictly positive weight
         w[0] = 0
         ajd_pham(covmats, sample_weight=w)
 
+
+def test_pham_zero_weight(get_covmats):
     # now test that setting one weight to almost zero it's almost
     # like not passing the matrix
-    V, D = ajd_pham(covmats[1:])
+    n_trials, n_channels, w_val = 5, 3, 2
+    covmats = get_covmats(n_trials, n_channels)
+    w = w_val * np.ones(n_trials)
+    V, D = ajd_pham(covmats[1:], sample_weight=w[1:])
     w[0] = 1e-12
 
     Vw, Dw = ajd_pham(covmats, sample_weight=w)
-    assert_allclose(V, Vw, rtol=1e-4, atol=1e-8)
-    assert_allclose(D, Dw[1:], rtol=1e-4, atol=1e-8)
+    assert V == approx(Vw, rel=1e-4, abs=1e-8)
+    assert D == approx(Dw[1:], rel=1e-4, abs=1e-8)
 
 
-def test_uwedge():
+@pytest.mark.parametrize("init", [True, False])
+def test_uwedge(init, get_covmats_params):
     """Test uwedge."""
-    covmats, diags, A = generate_cov(100, 3)
-    V, D = uwedge(covmats)
-    V, D = uwedge(covmats, init=A)
+    n_trials, n_channels = 5, 3
+    covmats, _, A = get_covmats_params(n_trials, n_channels)
+    if init:
+        V, D = uwedge(covmats)
+    else:
+        V, D = uwedge(covmats, init=A)
+    assert V.shape == (n_channels, n_channels)
+    assert D.shape == (n_trials, n_channels, n_channels)