Merge pull request #48 from orchardbirds/add_log_cosh

add log cosh error closes #47

bokbokbok/eval_metrics/regression/regression_eval_metrics.py

@@ -61,3 +61,31 @@ def root_mean_squared_log_error(yhat, dtrain, XGBoost=XGBoost):
             return 'RMSLE', float(np.sqrt(np.sum(elements) / len(y))), False
 
     return root_mean_squared_log_error
+
+
+def LogCoshMetric(XGBoost=False):
+    """
+    Calculates the Log Cosh Error
+    Args:
+        XGBoost (Bool): Set to True if using XGBoost. We assume LightGBM as default use.
+                        Note that you should also set `maximize=False` in the XGBoost train function
+
+    """
+    def log_cosh_error(yhat, dtrain, XGBoost=XGBoost):
+        """
+        Root Mean Squared Log Error.
+        All input labels are required to be greater than -1.
+
+        yhat: Predictions
+        dtrain: The XGBoost / LightGBM dataset
+        XGBoost (Bool): If XGBoost is to be implemented
+        """
+
+        y = dtrain.get_label()
+        elements = np.log(np.cosh(yhat - y))
+        if XGBoost:
+            return 'LogCosh', float(np.sum(elements) / len(y))
+        else:
+            return 'LogCosh', float(np.sum(elements) / len(y)), False
+
+    return log_cosh_error

bokbokbok/loss_functions/regression/regression_loss_functions.py

@@ -58,3 +58,59 @@ def squared_log_loss(
         return grad, hess
 
     return squared_log_loss
+
+
+def LogCoshLoss():
+    """
+    """
+
+    def _gradient(yhat, dtrain):
+        """Compute the log cosh gradient.
+
+        Args:
+            yhat (np.array): Predictions
+            dtrain: The XGBoost / LightGBM dataset
+
+        Returns:
+            log cosh gradient
+        """
+
+        y = dtrain.get_label()
+        return np.tanh(yhat - y)
+
+    def _hessian(yhat, dtrain):
+        """Compute the log cosh hessian.
+
+        Args:
+            yhat (np.array): Predictions
+            dtrain: The XGBoost / LightGBM dataset
+
+        Returns:
+            log cosh Hessian
+        """
+
+        y = dtrain.get_label()
+        return 1. / np.power(np.cosh(yhat - y), 2)
+
+    def log_cosh_loss(
+            yhat,
+            dtrain
+    ):
+        """
+        Calculate gradient and hessian for log cosh loss.
+
+        Args:
+            yhat (np.array): Predictions
+            dtrain: The XGBoost / LightGBM dataset
+
+        Returns:
+            grad: log cosh loss gradient
+            hess: log cosh loss Hessian
+        """
+        grad = _gradient(yhat, dtrain)
+
+        hess = _hessian(yhat, dtrain)
+
+        return grad, hess
+
+    return log_cosh_loss

docs/derivations/log_cosh.md

@@ -0,0 +1,13 @@
+## Log Cosh Error
+
+The equation for Log Cosh Error is:
+
+<img src="https://latex.codecogs.com/svg.latex?L_{LC}&space;=&space;\log(\cosh(\hat{y}&space;-&space;y))" title="L_{LC} = \log(\cosh(\hat{y} - y))" />
+
+We calculate the Gradient:
+
+<img src="https://latex.codecogs.com/svg.latex?G_{LC}&space;=&space;\tanh(\hat{y}&space;-&space;y)" title="G_{LC} = \tanh(\hat{y} - y)" />
+
+We also need to calculate the Hessian:
+
+<img src="https://latex.codecogs.com/svg.latex?H_{LC}&space;=&space;\frac{1}{\cosh^{2}(\hat{y}&space;-&space;y)}" title="H_{LC} = \frac{1}{\cosh^{2}(\hat{y} - y)}" />

docs/derivations/sle.md

@@ -1,6 +1,6 @@
 ## Squared Log Error
 
-The equations for Squared Log Error is:
+The equation for Squared Log Error is:
 
 <img src="https://latex.codecogs.com/svg.latex?L_{SLE}&space;=&space;\frac{1}{2}(\log(\hat{y}&space;&plus;&space;1)&space;-&space;\log(y&space;&plus;&space;1))^{2}" title="L_{SLE} = \frac{1}{2}(\log(\hat{y} + 1) - \log(y + 1))^{2}" />
 

mkdocs.yml

@@ -17,6 +17,7 @@ nav:
       - Weighted Cross Entropy: derivations/wce.md
       - Focal Loss: derivations/focal.md
       - Squared Log Error: derivations/sle.md
+      - Log Cosh Error: derivations/log_cosh.md
 
 plugins:
   - mkdocstrings: