TST: Test whether the positivity constraint works.

sklearn/decomposition/tests/test_dict_learning.py

@@ -34,6 +34,15 @@ def test_dict_learning_overcomplete():
     assert_true(dico.components_.shape == (n_components, n_features))
 
 
+def test_dict_learning_positivity():
+    n_components = 5
+    dico = DictionaryLearning(
+        n_components, transform_algorithm='lasso_lars', random_state=0,
+        positive=True).fit(X)
+    code = dico.transform(X)
+    assert_true((code >= 0).all())
+
+
 def test_dict_learning_reconstruction():
     n_components = 12
     dico = DictionaryLearning(n_components, transform_algorithm='omp',
@@ -111,6 +120,15 @@ def test_dict_learning_online_shapes():
     assert_equal(np.dot(code, dictionary).shape, X.shape)
 
 
+def test_dict_learning_online_positivity():
+    rng = np.random.RandomState(0)
+    n_components = 8
+    code, dictionary = dict_learning_online(X, n_components=n_components,
+                                            alpha=1, random_state=rng,
+                                            positive=True)
+    assert_true((code >= 0).all())
+
+
 def test_dict_learning_online_verbosity():
     n_components = 5
     # test verbosity
@@ -191,6 +209,20 @@ def test_sparse_encode_shapes():
         assert_equal(code.shape, (n_samples, n_components))
 
 
+def test_sparse_encode_positivity():
+    n_components = 12
+    rng = np.random.RandomState(0)
+    V = rng.randn(n_components, n_features)  # random init
+    V /= np.sum(V ** 2, axis=1)[:, np.newaxis]
+    for algo in ('lasso_lars', 'lasso_cd', 'lars', 'threshold'):
+        code = sparse_encode(X, V, algorithm=algo, positive=True)
+        assert_true((code >= 0).all())
+
+    assert_raises(
+        ValueError, sparse_encode, X, V, algorithm='omp', positive=True
+    )
+
+
 def test_sparse_encode_input():
     n_components = 100
     rng = np.random.RandomState(0)