Merge pull request #1073 from mglisse/tomato-kde-rescale

Rescale values for KDE to avoir underflow

src/python/gudhi/clustering/tomato.py

@@ -66,6 +66,7 @@ def __init__(
             graph_type (str): 'manual', 'knn' or 'radius'. Default is 'knn'.
             density_type (str): 'manual', 'DTM', 'logDTM', 'KDE' or 'logKDE'. When you have many points,
                 'KDE' and 'logKDE' tend to be slower. Default is 'logDTM'.
+                The values computed for 'DTM' or 'KDE' are not normalized (this does not affect the clustering).
             metric (str|Callable): metric used when calculating the distance between instances in a feature array.
                 Defaults to Minkowski of parameter p.
             kde_params (dict): if density_type is 'KDE' or 'logKDE', additional parameters passed directly to
@@ -224,6 +225,8 @@ def fit(self, X, y=None, weights=None):
 
             weights = KernelDensity(**kde_params).fit(self.points_).score_samples(self.points_)
             if self.density_type_ == "KDE":
+                # First rescale to avoid computing exp(-1000)
+                weights -= numpy.max(weights)
                 weights = numpy.exp(weights)
 
         # TODO: do it at the C++ level and/or in parallel if this is too slow?

src/python/test/test_tomato.py

@@ -63,3 +63,13 @@ def test_tomato_1():
     assert t.diagram_.size == 0
     assert t.max_weight_per_cc_.size == 1
     t.plot_diagram()
+
+
+def test_tomato_kde_underflow():
+    # 1D construction with 2 Gaussians, embedded in high dimension
+    X = np.zeros((200, 1000))
+    X[:100, 0] = np.random.default_rng().normal(-2, 1, 100)
+    X[100:, 0] = np.random.default_rng().normal(2, 1, 100)
+    # X[:,0].sort()
+    t = Tomato(density_type="KDE").fit(X)
+    assert (t.weights_ != 0).all()