Add Learner2D loss function 'thresholded_loss_factory'

Author: basnijholt

adaptive/learner/learner2D.py

@@ -231,6 +231,64 @@ def default_loss(ip: LinearNDInterpolator) -> np.ndarray:
     return losses
 
 
+def thresholded_loss_factory(
+    lower_threshold: float | None = None,
+    upper_threshold: float | None = None,
+    priority_factor: float = 0.1,
+) -> Callable[[LinearNDInterpolator], np.ndarray]:
+    """
+    Factory function to create a custom loss function that deprioritizes
+    values above an upper threshold and below a lower threshold.
+
+    Parameters
+    ----------
+    lower_threshold : float, optional
+        The lower threshold for deprioritizing values. If None (default),
+        there is no lower threshold.
+    upper_threshold : float, optional
+        The upper threshold for deprioritizing values. If None (default),
+        there is no upper threshold.
+    priority_factor : float, default: 0.1
+        The factor by which the loss is multiplied for values outside
+        the specified thresholds.
+
+    Returns
+    -------
+    custom_loss : Callable[[LinearNDInterpolator], np.ndarray]
+        A custom loss function that can be used with Learner2D.
+    """
+
+    def custom_loss(ip: LinearNDInterpolator) -> np.ndarray:
+        """Loss function that deprioritizes values above an upper and lower threshold.
+
+        Parameters
+        ----------
+        ip : `scipy.interpolate.LinearNDInterpolator` instance
+
+        Returns
+        -------
+        losses : numpy.ndarray
+            Loss per triangle in ``ip.tri``.
+        """
+        losses = default_loss(ip)
+
+        if lower_threshold is not None or upper_threshold is not None:
+            values = ip.values
+            if lower_threshold is not None:
+                mask_lower = values < lower_threshold
+                if mask_lower.any():
+                    losses[mask_lower] *= priority_factor
+
+            if upper_threshold is not None:
+                mask_upper = values > upper_threshold
+                if mask_upper.any():
+                    losses[mask_upper] *= priority_factor
+
+        return losses
+
+    return custom_loss
+
+
 def choose_point_in_triangle(triangle: np.ndarray, max_badness: int) -> np.ndarray:
     """Choose a new point in inside a triangle.