Some robustifying fixes on Atol fit and tests for vectorizers

src/python/gudhi/representations/vector_methods.py

@@ -807,20 +807,28 @@ def fit(self, X, y=None, sample_weight=None):
             self
         """
         if not hasattr(self.quantiser, 'fit'):
-            raise TypeError("quantiser %s has no `fit` attribute." % (self.quantiser))
-
-        # In fitting we remove infinite death time points so that every center is finite
-        X = [dgm[~np.isinf(dgm).any(axis=1), :] for dgm in X]
+            raise TypeError(f"quantiser {self.quantiser} has no `fit` attribute.")
 
         if sample_weight is None:
             sample_weight = [self.get_weighting_method()(measure) for measure in X]
 
         measures_concat = np.concatenate(X)
         weights_concat = np.concatenate(sample_weight)
-
-        self.quantiser.fit(X=measures_concat, sample_weight=weights_concat)
-
+        # In fitting we remove infinite birth/death time points so that every center is finite
+        filtered_measures_concat = measures_concat[~np.isinf(measures_concat).any(axis=1), :] if len(measures_concat) else measures_concat
+        filtered_weights_concat = weights_concat[~np.isinf(measures_concat).any(axis=1)] if len(measures_concat) else weights_concat
+
+        n_clusters = self.quantiser.n_clusters
+        n_points = len(filtered_measures_concat)
+        if n_points < n_clusters:
+            # If not enough points to fit (including 0), let's arbitrarily put centers in [0, 1)^2
+            print(f"[Atol] had {n_points} points to fit {n_clusters} clusters, adding random points in [0, 1)^2.")
+            filtered_weights_concat = np.concatenate((filtered_weights_concat, np.ones(shape=(n_clusters - n_points))))
+            filtered_measures_concat = np.concatenate((filtered_measures_concat, np.random.random((n_clusters - n_points, 2))))
+
+        self.quantiser.fit(X=filtered_measures_concat, sample_weight=filtered_weights_concat)
         self.centers = self.quantiser.cluster_centers_
+
         # Hack, but some people are unhappy if the order depends on the version of sklearn
         self.centers = self.centers[np.lexsort(self.centers.T)]
         if self.quantiser.n_clusters == 1:

src/python/test/test_representations_interface.py

@@ -0,0 +1,90 @@
+from copy import deepcopy
+import numpy as np
+
+from sklearn.cluster import KMeans
+
+from gudhi.representations import (Atol, Landscape, Silhouette, BettiCurve, ComplexPolynomial, \
+                                   TopologicalVector, PersistenceImage, Entropy)
+
+vectorizers = {
+    "atol": Atol(quantiser=KMeans(n_clusters=2, random_state=202312, n_init="auto")),
+    # "betti": BettiCurve(),
+}
+
+diag1 = [np.array([[0., np.inf],
+                   [0., 8.94427191],
+                   [0., 7.28010989],
+                   [0., 6.08276253],
+                   [0., 5.83095189],
+                   [0., 5.38516481],
+                   [0., 5.]]),
+         np.array([[11., np.inf],
+                   [6.32455532, 6.70820393]]),
+         np.empty(shape=[0, 2])]
+
+diag2 = [np.array([[0., np.inf],
+                   [0., 8.94427191],
+                   [0., 7.28010989],
+                   [0., 6.08276253],
+                   [0., 5.83095189],
+                   [0., 5.38516481],
+                   [0., 5.]]),
+         np.array([[11., np.inf],
+                   [6.32455532, 6.70820393]]),
+         np.array([[0., np.inf],
+                   [0., 1]])]
+
+diag3 = [np.empty(shape=[0, 2])]
+
+
+def test_fit():
+    print(f" > Testing `fit`.")
+    for name, vectorizer in vectorizers.items():
+        print(f" >> Testing {name}")
+        deepcopy(vectorizer).fit(X=[diag1[0], diag2[0]])
+
+
+def test_fit_empty():
+    print(f" > Testing `fit_empty`.")
+    for name, vectorizer in vectorizers.items():
+        print(f" >> Testing {name}")
+        deepcopy(vectorizer).fit(X=[diag3[0], diag3[0]])
+
+
+def test_transform():
+    print(f" > Testing `transform`.")
+    for name, vectorizer in vectorizers.items():
+        print(f" >> Testing {name}")
+        deepcopy(vectorizer).fit_transform(X=[diag1[0], diag2[0], diag3[0]])
+
+
+def test_transform_empty():
+    print(f" > Testing `transform_empty`.")
+    for name, vectorizer in vectorizers.items():
+        print(f" >> Testing {name}")
+        copy_vec = deepcopy(vectorizer).fit(X=[diag1[0], diag2[0]])
+        copy_vec.transform(X=[diag3[0], diag3[0]])
+
+
+def test_set_output():
+    print(f" > Testing `set_output`.")
+    try:
+        import pandas
+        for name, vectorizer in vectorizers.items():
+            print(f" >> Testing {name}")
+            deepcopy(vectorizer).set_output(transform="pandas")
+    except ImportError:
+        print("Missing pandas, skipping set_output test")
+
+
+def test_compose():
+    print(f" > Testing composition with `sklearn.compose.ColumnTransformer`.")
+    from sklearn.compose import ColumnTransformer
+    for name, vectorizer in vectorizers.items():
+        print(f" >> Testing {name}")
+        ct = ColumnTransformer([
+            (f"{name}-0", deepcopy(vectorizer), 0),
+            (f"{name}-1", deepcopy(vectorizer), 1),
+            (f"{name}-2", deepcopy(vectorizer), 2)]
+        )
+        ct.fit_transform(X=[diag1, diag2])