Fix get_decimal_precisions IndexError on scientific-notation modes (#442)

dice_ml/data_interfaces/public_data_interface.py

@@ -381,6 +381,45 @@ def from_dummies(self, data, prefix_sep='_'):
             out.drop(cols, axis=1, inplace=True)
         return out
 
+    @staticmethod
+    def _decimal_precision_of(value):
+        """Number of decimal digits in the default repr of *value*.
+
+        The historical implementation was ``len(str(value).split('.')[1])``,
+        which mirrors how many digits Python prints after the decimal point
+        in the default repr. That works for ordinary floats (``'0.25'`` → 2)
+        but IndexErrors when ``str(value)`` renders in scientific notation —
+        ``str(np.float32(1e-6))`` is ``'1e-06'`` with no ``'.'`` (issue #442).
+
+        This helper preserves the original semantics for the common case and
+        re-derives an equivalent count from the scientific-notation form,
+        rather than from a re-formatted ``%.20f`` string (which leaks
+        float64 ↔ float32 round-off digits and inflates precision).
+
+        Examples:
+            ``0.25``       → 2
+            ``1e-06``      → 6  (no mantissa fraction, exponent -6)
+            ``2.5e-3``     → 4  (mantissa '2.5' has 1 decimal, exponent -3)
+            ``1e+16``      → 0  (exponent positive ⇒ no decimal digits)
+        """
+        s = str(value).lower()
+        if 'e' not in s:
+            # Plain decimal repr, e.g. '0.25' or '17'.
+            if '.' not in s:
+                return 0
+            return len(s.split('.')[1])
+        # Scientific notation, e.g. '1e-06', '2.5e-3', '1e+16'.
+        mantissa, _, exponent = s.partition('e')
+        mantissa_decimals = len(mantissa.split('.')[1]) if '.' in mantissa else 0
+        try:
+            exp_value = int(exponent)
+        except ValueError:
+            return mantissa_decimals
+        # Decimal places needed = mantissa decimals minus the exponent shift.
+        # A negative exponent moves digits right; a positive one moves them
+        # left (and can cancel mantissa decimals out entirely).
+        return max(mantissa_decimals - exp_value, 0)
+
     def get_decimal_precisions(self, output_type="list"):
         """"Gets the precision of continuous features in the data."""
         # if the precision of a continuous feature is not given, we use the maximum precision of the modes to capture the
@@ -393,9 +432,12 @@ def get_decimal_precisions(self, output_type="list"):
                 precisions_dict[col] = self.continuous_features_precision[col]
             elif self.data_df[col].dtype == np.float32 or self.data_df[col].dtype == np.float64:
                 modes = self.data_df[col].mode()
-                maxp = len(str(modes[0]).split('.')[1])  # maxp stores the maximum precision of the modes
+                # Use _decimal_precision_of instead of raw str().split('.')[1]
+                # because the latter IndexErrors when the mode renders in
+                # scientific notation (e.g. 1e-06, 1e+16) — see #442.
+                maxp = self._decimal_precision_of(modes[0])
                 for mx in range(len(modes)):
-                    prec = len(str(modes[mx]).split('.')[1])
+                    prec = self._decimal_precision_of(modes[mx])
                     if prec > maxp:
                         maxp = prec
                 precisions[ix] = maxp

tests/test_data_interface/test_public_data_interface.py

@@ -38,6 +38,41 @@ def test_feature_precision(self):
         assert self.d.get_decimal_precisions()[1] == 2
 
 
+class TestDecimalPrecisionsScientificNotation:
+    """Regression for issue #442.
+
+    `str(np.float64(1e-6))` is `'1e-06'`; the historical
+    `str(value).split('.')[1]` then IndexErrors on mode values whose default
+    repr is in scientific notation (very small, e.g. <=1e-5, or very large,
+    e.g. >=1e16). This makes DiCE's Data() constructor blow up on legitimate
+    columns of currency / probability / nano-second magnitudes.
+    """
+
+    @pytest.mark.parametrize(
+        "values, expected_precision",
+        [
+            # mode renders as '1e-06' — used to IndexError, now returns 6.
+            ([1e-6, 1e-6, 1e-6], 6),
+            # mode renders as '2.5e-3' — mantissa has 1 decimal, exponent -3 ⇒ 4.
+            ([2.5e-3, 2.5e-3, 2.5e-3], 4),
+            # mode renders as '1e+16' — positive exponent collapses to 0.
+            ([1e16, 1e16, 1e16], 0),
+            # ordinary floats still work — mode 0.5 has 1 decimal.
+            ([0.5, 0.5, 0.25], 1),
+        ],
+    )
+    def test_get_decimal_precisions_handles_scientific_notation(
+        self, values, expected_precision
+    ):
+        df = pd.DataFrame({"feat": values, "outcome": [0, 1, 0]})
+        d = dice_ml.Data(
+            dataframe=df, continuous_features=["feat"], outcome_name="outcome"
+        )
+        # Must not raise; must return a sensible precision.
+        precisions = d.get_decimal_precisions()
+        assert precisions[0] == expected_precision
+
+
 class DataTypeCombinations(Enum):
     Incorrect = 0
     AsNone = 1