xgboost: add clipping of loss values to the float32 limits

Internally the XGBoost routine operates a conversion to float32, and this fails
if the absolute values of the loss are exceedingly large. This commit adds a
simple clipping operation before the loss values are passed to the XGBoost
package.

Author: AldoGl

.pylintrc

@@ -11,6 +11,7 @@ disable=C0103,C0301,R0801,W1514,W1505
 
 [IMPORTS]
 ignored-modules=
+known-third-party=xgboost
 
 [DESIGN]
 # min-public-methods=1

black_it/samplers/xgboost.py

@@ -15,6 +15,7 @@
 # along with this program. If not, see <http://www.gnu.org/licenses/>.
 
 """This module contains the implementation of the XGBoost sampling."""
+import warnings
 from typing import Optional, cast
 
 import numpy as np
@@ -23,6 +24,10 @@
 
 from black_it.samplers.surrogate import MLSurrogateSampler
 
+MAX_FLOAT32 = np.finfo(np.float32).max
+MIN_FLOAT32 = np.finfo(np.float32).min
+EPS_FLOAT32 = np.finfo(np.float32).eps
+
 
 class XGBoostSampler(MLSurrogateSampler):
     """This class implements xgboost sampling."""
@@ -95,9 +100,31 @@ def n_estimators(self) -> int:
         """Get the number of estimators."""
         return self._n_estimators
 
+    @staticmethod
+    def _clip_losses(y: NDArray[np.float64]) -> NDArray[np.float64]:
+        """Check that loss values fall within the float32 limits needed for XGBoost to work."""
+        large_floats = np.where(y >= MAX_FLOAT32)
+        small_floats = np.where(y <= MIN_FLOAT32)
+
+        if len(large_floats) == 0 and len(small_floats) == 0:
+            return y
+
+        warnings.warn(
+            "Found loss values out of float32 limits, clipping them for XGBoost.",
+            RuntimeWarning,
+        )
+        if len(large_floats) > 0:
+            y[large_floats] = MAX_FLOAT32 - EPS_FLOAT32
+
+        if len(small_floats) > 0:
+            y[small_floats] = MIN_FLOAT32 + EPS_FLOAT32
+
+        return y
+
     def fit(self, X: NDArray[np.float64], y: NDArray[np.float64]) -> None:
         """Fit a xgboost surrogate model."""
         # prepare data
+        y = self._clip_losses(y)  # pylint: disable=W0212
         _ = xgb.DMatrix(data=X, label=y)
 
         # train surrogate

setup.cfg

@@ -35,6 +35,7 @@ force_grid_wrap=0
 use_parentheses=True
 line_length=88
 ensure_newline_before_comments=True
+known_third_party=xgboost
 
 [black]
 exclude = "scripts/whitelists/"

tests/test_samplers/test_xgboost.py

@@ -15,8 +15,6 @@
 # along with this program. If not, see <http://www.gnu.org/licenses/>.
 """This module contains tests for the xgboost sampler."""
 import numpy as np
-import pytest
-import xgboost as xgb
 
 from black_it.calibrator import Calibrator
 from black_it.loss_functions.msm import MethodOfMomentsLoss
@@ -28,6 +26,10 @@
 
 expected_params = np.array([[0.24, 0.26], [0.26, 0.02], [0.08, 0.24], [0.15, 0.15]])
 
+MAX_FLOAT32 = np.finfo(np.float32).max
+MIN_FLOAT32 = np.finfo(np.float32).min
+EPS_FLOAT32 = np.finfo(np.float32).eps
+
 
 def test_xgboost_2d() -> None:
     """Test the xgboost sampler, 2d."""
@@ -94,10 +96,28 @@ def test_clip_losses() -> None:
         random_state=0,
     )
 
-    # the calibration breaks due to losses exceeding the limits of float32
+    # verify that the calibration does not break,
+    # it would without the call to _clip_losses
+    _, losses = cal.calibrate(1)
+
+    assert np.allclose(
+        losses,
+        np.array(
+            [
+                7.46098998e02,
+                5.80544566e17,
+                3.40282347e38,
+                3.40282347e38,
+                3.40282347e38,
+                2.94273501e41,
+            ]
+        ),
+    )
+
+    # verify that _clip_losses works as expected
+    y = np.array([0.0, -1e40, 1e40])
+    y2 = xgboost._clip_losses(y)  # pylint: disable=W0212
 
-    with pytest.raises(
-        xgb.core.XGBoostError,
-        match=r"Label contains NaN, infinity or a value too large",
-    ):
-        _, losses = cal.calibrate(1)
+    assert (
+        y2 == np.array([0.0, MIN_FLOAT32 + EPS_FLOAT32, MAX_FLOAT32 - EPS_FLOAT32])
+    ).all()