Update to version 0.4

CHANGELOG.md

@@ -1,5 +1,11 @@
 # Changelog
 
+## [0.4] - 2026-01-05
+
+### Fixed
+
+- `build_reliability_diagram`: Error when scores are distributed uniformly.
+
 ## [0.3] - 2025-12-26
 
 ### Fixed
@@ -17,7 +23,8 @@
 - `calibration_error` function
 - `build_reliability_diagram` function
 
-[Unreleased]: https://github.com/yourusername/yourproject/compare/0.3...HEAD
-[0.3]: https://github.com/yourusername/yourproject/compare/0.2...0.3
-[0.2]: https://github.com/yourusername/yourproject/compare/0.1...0.2
-[0.1]: https://github.com/yourusername/yourproject/releases/tag/0.1
+[Unreleased]: https://github.com/khiopslab/khalib/compare/0.4...HEAD
+[0.4]: https://github.com/khiopslab/khalib/compare/0.3...0.4
+[0.3]: https://github.com/khiopslab/khalib/compare/0.2...0.3
+[0.2]: https://github.com/khiopslab/khalib/compare/0.1...0.2
+[0.1]: https://github.com/khiopslab/khalib/releases/tag/0.1

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "khalib"
-version = "0.3"
+version = "0.4"
 description = "Classifier Calibration with Khiops"
 authors = [{ name = "Felipe Olmos", email = "[email protected]" }]
 requires-python = ">=3.11"

src/khalib/main.py

@@ -233,9 +233,6 @@ def from_data(
                 do_data_preparation_only=True,
             )
             results = kh.read_analysis_results_file(f"{work_dir}/report.khj")
-            # import shutil
-
-            # shutil.copy(f"{work_dir}/report.khj", "tmp.khj")
 
         # Initialize the histogram
         if y is not None:
@@ -287,10 +284,13 @@ def from_data(
                 else:
                     if use_finest:
                         histogram_index = -1
-                    else:
+                    elif 100 in score_stats.modl_histograms.information_rates:
                         histogram_index = (
                             score_stats.modl_histograms.information_rates.index(100)
                         )
+                    else:
+                        assert len(score_stats.modl_histograms.information_rates) == 1
+                        histogram_index = 0
                     breakpoints = score_stats.modl_histograms.histograms[
                         histogram_index
                     ].bounds
@@ -954,31 +954,11 @@ def build_reliability_diagram(
     fig.subplots_adjust(hspace=0)
     fig.suptitle("Reliability Diagram")
 
-    # Build a unsupervised histogram to to detect the dirac masses case
+    # Build a unsupervised histogram
     uhist_y_scores = Histogram.from_data(y_scores, use_finest=True)
-    dirac_indexes = []
-    if uhist_y_scores.freqs[1] == 0 and (
-        uhist_y_scores.bins[0][1] - uhist_y_scores.bins[0][0] < dirac_threshold
-    ):
-        dirac_indexes.append(True)
-    else:
-        dirac_indexes.append(False)
-    for i in range(1, uhist_y_scores.n_bins - 1):
-        cur_left, cur_right = uhist_y_scores.bins[i]
-        if (
-            uhist_y_scores.freqs[i - 1] == 0
-            and uhist_y_scores.freqs[i + 1] == 0
-            and (cur_right - cur_left < dirac_threshold)
-        ):
-            dirac_indexes.append(True)
-        else:
-            dirac_indexes.append(False)
-    if uhist_y_scores.freqs[-2] == 0 and (
-        uhist_y_scores.bins[-1][1] - uhist_y_scores.bins[-1][0] < dirac_threshold
-    ):
-        dirac_indexes.append(True)
-    else:
-        dirac_indexes.append(False)
+
+    # Compute the dirac mass indexes
+    dirac_indexes = compute_dirac_indexes(uhist_y_scores, dirac_threshold)
 
     # Compute the supervised score histogram
     hist_y_scores = Histogram.from_data(y_scores, y)
@@ -1091,3 +1071,43 @@ def build_reliability_diagram(
             )
 
     return fig, axs
+
+
+def compute_dirac_indexes(uhist, dirac_threshold):
+    """Computes the dirac mass indexes of a histogram
+
+    We declare a dirac mass bin if:
+
+    - it is surrounded by empty bins.
+    - its length is less than ``dirac_threshold``
+
+    """
+    dirac_indexes = []
+    if (
+        len(uhist.freqs) > 1
+        and uhist.freqs[1] == 0
+        and (uhist.bins[0][1] - uhist.bins[0][0] < dirac_threshold)
+    ):
+        dirac_indexes.append(True)
+    else:
+        dirac_indexes.append(False)
+    for i in range(1, uhist.n_bins - 1):
+        cur_left, cur_right = uhist.bins[i]
+        if (
+            uhist.freqs[i - 1] == 0
+            and uhist.freqs[i + 1] == 0
+            and (cur_right - cur_left < dirac_threshold)
+        ):
+            dirac_indexes.append(True)
+        else:
+            dirac_indexes.append(False)
+    if (
+        len(uhist.freqs) > 2
+        and uhist.freqs[-2] == 0
+        and (uhist.bins[-1][1] - uhist.bins[-1][0] < dirac_threshold)
+    ):
+        dirac_indexes.append(True)
+    else:
+        dirac_indexes.append(False)
+
+    return dirac_indexes

src/tests/test_all.py

@@ -11,7 +11,8 @@
 from sklearn.svm import SVC
 from sklearn.utils.validation import check_is_fitted
 
-from khalib import Histogram, KhalibClassifier, calibration_error
+from khalib import Histogram, KhalibClassifier, calibrate_binary, calibration_error
+from khalib.main import compute_dirac_indexes
 
 
 @pytest.fixture(name="data_root_dir")
@@ -443,3 +444,29 @@ def test_calibrator(
                 y_scores_calib_test, y_test, multi_class_method="top-label"
             )
             assert ece == pytest.approx(expected_ece, rel=1e-2)
+
+
+class TestDiracHeuristic:
+    def test_uniform(self):
+        # The uniform distribution should have only one bin and no diracs
+        rng = np.random.default_rng(seed=1234567)
+        y_scores = rng.uniform(size=2000)
+        uhist = Histogram.from_data(y_scores, use_finest=True)
+        assert compute_dirac_indexes(uhist, 1e-06) == [False, False]
+
+    def test_dirac(self):
+        y_scores = np.array([0.2] * 250 + [0.5] * 250 + [0.9] * 500)
+        y = np.array(
+            [0] * 200 + [1] * 50 + [0] * 125 + [1] * 125 + [0] * 50 + [1] * 450
+        )
+        hist = Histogram.from_data(y_scores, y=y)
+        y_scores_calib = calibrate_binary(y_scores, hist, only_positive=True)
+        uhist = Histogram.from_data(y_scores_calib, use_finest=True)
+        assert compute_dirac_indexes(uhist, 1e-06) == [True, False, True, False, True]
+
+    def test_beta(self):
+        # A sufficiently large continuous distribution sample should have no diracs
+        rng = np.random.default_rng(seed=1234567)
+        y_scores = rng.beta(a=0.5, b=0.5, size=2000)
+        uhist = Histogram.from_data(y_scores, use_finest=True)
+        assert not all(compute_dirac_indexes(uhist, 1e-06))