Fixes for RSL, plus more tests

Author: lmcinnes

hdbscan/robust_single_linkage_.py

@@ -28,11 +28,8 @@
 FAST_METRICS = KDTree.valid_metrics + BallTree.valid_metrics
 
 
-def _rsl_generic(X, k=5, alpha=1.4142135623730951, metric='minkowski', p=2):
-    if metric == 'minkowski':
-        distance_matrix = pairwise_distances(X, metric=metric, p=p)
-    else:
-        distance_matrix = pairwise_distances(X, metric=metric)
+def _rsl_generic(X, k=5, alpha=1.4142135623730951, metric='euclidean', **kwargs):
+    distance_matrix = pairwise_distances(X, metric=metric, **kwargs)
 
     mutual_reachability_ = mutual_reachability(distance_matrix, k)
 
@@ -45,7 +42,7 @@ def _rsl_generic(X, k=5, alpha=1.4142135623730951, metric='minkowski', p=2):
     return single_linkage_tree
 
 
-def _rsl_prims_kdtree(X, k=5, alpha=1.4142135623730951, metric='minkowski', **kwargs):
+def _rsl_prims_kdtree(X, k=5, alpha=1.4142135623730951, metric='euclidean', **kwargs):
 
     # The Cython routines used require contiguous arrays
     if not X.flags['C_CONTIGUOUS']:
@@ -67,7 +64,7 @@ def _rsl_prims_kdtree(X, k=5, alpha=1.4142135623730951, metric='minkowski', **kw
     return single_linkage_tree
 
 
-def _rsl_prims_balltree(X, k=5, alpha=1.4142135623730951, metric='minkowski', **kwargs):
+def _rsl_prims_balltree(X, k=5, alpha=1.4142135623730951, metric='euclidean', **kwargs):
 
     # The Cython routines used require contiguous arrays
     if not X.flags['C_CONTIGUOUS']:
@@ -90,7 +87,7 @@ def _rsl_prims_balltree(X, k=5, alpha=1.4142135623730951, metric='minkowski', **
 
 
 def _rsl_boruvka_kdtree(X, k=5, alpha=1.0,
-                        metric='minkowski', leaf_size=40, **kwargs):
+                        metric='euclidean', leaf_size=40, **kwargs):
 
     dim = X.shape[0]
     min_samples = min(dim - 1, k)
@@ -107,7 +104,7 @@ def _rsl_boruvka_kdtree(X, k=5, alpha=1.0,
 
 
 def _rsl_boruvka_balltree(X, k=5, alpha=1.0,
-                          metric='minkowski', leaf_size=40, **kwargs):
+                          metric='euclidean', leaf_size=40, **kwargs):
 
     dim = X.shape[0]
     min_samples = min(dim - 1, k)
@@ -342,7 +339,7 @@ class RobustSingleLinkage(BaseEstimator, ClusterMixin):
     """
 
     def __init__(self, cut=0.4, k=5, alpha=1.4142135623730951, gamma=5, metric='euclidean',
-                 algorithm='best'):
+                 algorithm='best', **kwargs):
 
         self.cut = cut
         self.k = k
@@ -351,6 +348,8 @@ def __init__(self, cut=0.4, k=5, alpha=1.4142135623730951, gamma=5, metric='eucl
         self.metric = metric
         self.algorithm = algorithm
 
+        self._metric_kwargs = kwargs
+
         self._cluster_hierarchy_ = None
 
     def fit(self, X, y=None):
@@ -364,7 +363,11 @@ def fit(self, X, y=None):
             ``metric='precomputed'``.
         """
         X = check_array(X, accept_sparse='csr')
-        self.labels_, self._cluster_hierarchy_ = robust_single_linkage(X, **self.get_params())
+
+        kwargs = self.get_params()
+        kwargs.update(self._metric_kwargs)
+
+        self.labels_, self._cluster_hierarchy_ = robust_single_linkage(X, **kwargs)
         return self
 
     def fit_predict(self, X, y=None):

hdbscan/tests/test_rsl.py

@@ -19,6 +19,7 @@
 # from sklearn.cluster.tests.common import generate_clustered_data
 
 from sklearn import datasets
+import warnings
 
 from sklearn.datasets import make_blobs
 from sklearn.utils import shuffle
@@ -96,6 +97,31 @@ def test_rsl_prims_kdtree():
     n_clusters_2 = len(set(labels)) - int(-1 in labels)
     assert_equal(n_clusters_2, n_clusters)
 
+def test_rsl_unavailable_hierarchy():
+    clusterer = RobustSingleLinkage()
+    with warnings.catch_warnings(record=True) as w:
+        tree = clusterer.cluster_hierarchy_
+        assert(len(w) > 0)
+        assert(tree is None)
+
+def test_rsl_hierarchy():
+    clusterer = RobustSingleLinkage().fit(X)
+    assert(clusterer.cluster_hierarchy_ is not None)
+
+def test_rsl_high_dimensional():
+    H, y = make_blobs(n_samples=50, random_state=0, n_features=64)
+    # H, y = shuffle(X, y, random_state=7)
+    H = StandardScaler().fit_transform(H)
+    labels, tree = robust_single_linkage(H, 5.5)
+    n_clusters_1 = len(set(labels)) - int(-1 in labels)
+    print(n_clusters_1)
+    assert_equal(n_clusters_1, n_clusters)
+
+    labels = RobustSingleLinkage(cut=5.5, algorithm='best', metric='seuclidean', V=np.ones(H.shape[1])).fit(H).labels_
+    n_clusters_2 = len(set(labels)) - int(-1 in labels)
+    print n_clusters_2
+    assert_equal(n_clusters_2, n_clusters)
+
 def test_rsl_badargs():
     assert_raises(ValueError,
                   robust_single_linkage,
@@ -154,6 +180,9 @@ def test_rsl_badargs():
     assert_raises(ValueError,
                   robust_single_linkage,
                   X, 0.4, leaf_size=0)
+    assert_raises(ValueError,
+                  robust_single_linkage,
+                  X, 0.4, gamma=0)
 
 def test_rsl_is_sklearn_estimator():