Feat add new loss functions

pydeepflow/device.py

@@ -194,6 +194,22 @@ def tanh(self, x):
         """
         return cp.tanh(x) if self.use_gpu else np.tanh(x)
 
+    def cosh(self, x):
+        """
+        Computes the hyperbolic cosine of the input array.
+
+        Parameters:
+        -----------
+        x : np.ndarray or cp.ndarray
+            The input array.
+
+        Returns:
+        --------
+        np.ndarray or cp.ndarray
+            The hyperbolic cosine of the input.
+        """
+        return cp.cosh(x) if self.use_gpu else np.cosh(x)
+
     def sum(self, x, axis=None, keepdims=False):
         """
         Sums the elements of an array along a specified axis.

pydeepflow/losses.py

@@ -212,6 +212,48 @@ def huber_loss_derivative(y_true, y_pred, device, delta=1.0):
     is_small_error = device.abs(error) <= delta
     return device.where(is_small_error, error, delta * device.sign(error))
 
+def log_cosh_loss(y_true, y_pred, device):
+    """
+    Computes the Log-Cosh loss for regression tasks.
+
+    Parameters:
+    -----------
+    y_true : np.ndarray or cp.ndarray
+        Ground truth values.
+    y_pred : np.ndarray or cp.ndarray
+        Predicted values.
+    device : Device
+        The device instance (CPU or GPU) to perform calculations.
+
+    Returns:
+    --------
+    float
+        The Log-Cosh loss.
+    """
+    error = y_pred - y_true
+    return device.mean(device.log(device.cosh(error)))
+
+def log_cosh_loss_derivative(y_true, y_pred, device):
+    """
+    Computes the derivative of the Log-Cosh loss.
+
+    Parameters:
+    -----------
+    y_true : np.ndarray or cp.ndarray
+        Ground truth values.
+    y_pred : np.ndarray or cp.ndarray
+        Predicted values.
+    device : Device
+        The device instance (CPU or GPU) to perform calculations.
+
+    Returns:
+    --------
+    np.ndarray or cp.ndarray
+        The derivative of the Log-Cosh loss with respect to predictions.
+    """
+    error = y_pred - y_true
+    return device.tanh(error)
+
 # Get the appropriate loss function
 def get_loss_function(loss_name):
     """
@@ -242,6 +284,8 @@ def get_loss_function(loss_name):
         return hinge_loss
     elif loss_name == 'huber':
         return huber_loss
+    elif loss_name == 'log_cosh':
+        return log_cosh_loss
     else:
         raise ValueError(f"Unsupported loss function: {loss_name}")
 
@@ -275,5 +319,7 @@ def get_loss_derivative(loss_name):
         return hinge_loss_derivative
     elif loss_name == 'huber':
         return huber_loss_derivative
+    elif loss_name == 'log_cosh':
+        return log_cosh_loss_derivative
     else:
         raise ValueError(f"Unsupported loss derivative: {loss_name}")

tests/test_losses.py

@@ -5,7 +5,8 @@
     binary_crossentropy, mse, mse_derivative,
     categorical_crossentropy, categorical_crossentropy_derivative,
     hinge_loss, hinge_loss_derivative,
-    huber_loss, huber_loss_derivative
+    huber_loss, huber_loss_derivative,
+    log_cosh_loss, log_cosh_loss_derivative
 )
 from pydeepflow.device import Device
 
@@ -85,6 +86,22 @@ def test_huber_loss_derivative(self):
         result = huber_loss_derivative(y_true, y_pred, self.device_cpu, delta)
         np.testing.assert_array_almost_equal(result, expected)
 
+    def test_log_cosh_loss(self):
+        y_true = np.array([1, 2, 3])
+        y_pred = np.array([2, 2, 4])
+        error = y_pred - y_true
+        expected = np.mean(np.log(np.cosh(error)))
+        result = log_cosh_loss(y_true, y_pred, self.device_cpu)
+        self.assertAlmostEqual(result, expected)
+
+    def test_log_cosh_loss_derivative(self):
+        y_true = np.array([1.0, 2.0, 3.0])
+        y_pred = np.array([1.5, 1.7, 2.5])
+        error = y_pred - y_true
+        expected = np.tanh(error)
+        result = log_cosh_loss_derivative(y_true, y_pred, self.device_cpu)
+        np.testing.assert_array_almost_equal(result, expected)
+
 if __name__ == "__main__":
     unittest.main()