feat : Implement BCELoss

trueml/losses/__init__.py

@@ -1,7 +1,8 @@
 from .mean_absolute_error import MAEloss
 from .mean_squared_error import MSEloss
-
+from .binary_cross_entropy_loss import BCELoss
 __all__ = [
     "MAEloss",
     "MSEloss",
+    "BCELoss"
 ]

trueml/losses/binary_cross_entropy_loss.py

@@ -0,0 +1,49 @@
+import numpy as np
+
+class BCELoss:
+    """
+    Compute the binary cross-entropy (BCE) loss.
+
+    Formula:
+        -mean(y_true * log(y_pred) + (1 - y_true) * log(1 - y_pred))
+
+    BCE measures the performance of a classification model whose output is a probability value between 0 and 1.
+
+    Examples
+    --------
+    >>> bce = BCELoss()
+    >>> bce([1, 0], [0.9, 0.1])
+    0.10536051565782628
+    """
+
+    @staticmethod
+    def surface(y_true, y_pred):
+        # Clip to prevent log(0) causing NaN/inf values
+        eps = 1e-15
+        y_pred = np.clip(y_pred, eps, 1 - eps)
+        return -(y_true * np.log(y_pred) + (1 - y_true) * np.log(1 - y_pred))
+
+    def __call__(self, y_true, y_pred):
+        y_true = np.array(y_true)
+        y_pred = np.array(y_pred)
+        return np.mean(self.surface(y_true, y_pred))
+
+    def grad(self, y_true, y_pred):
+        """
+        Compute gradient of BCE w.r.t predictions.
+
+        Returns:
+            array-like
+                Gradient of the loss with respect to predictions (dL/dy_pred)
+        """
+        y_true = np.array(y_true)
+        y_pred = np.array(y_pred)
+
+        # Clip to prevent division by zero in gradients
+        eps = 1e-15
+        y_pred = np.clip(y_pred, eps, 1 - eps)
+
+        n = y_true.shape[0]
+
+        # Standard analytical derivative of BCE divided by n for the mean reduction
+        return (1 / n) * ((y_pred - y_true) / (y_pred * (1 - y_pred)))