DropLR.py

class DropLR:
    """
    Custom learning rate scheduler that reduces learning rate when validation accuracy plateaus.
    Monitors validation accuracy and signals when to drop learning rate based on lack of improvement.
    """
    def __init__(self, tolerance=7, delta_factor=1e-3):
        """
        Initialize the DropLR scheduler.
        
        Args:
            tolerance (int): Number of epochs to wait without improvement before dropping LR
            delta_factor (float): Minimum relative improvement threshold (e.g., 1e-3 = 0.1% improvement)
        """
        self.tolerance = tolerance
        self.delta_factor = delta_factor
        self.counter = 0
        self.max_validation_accuracy = -float('inf')
        self.drop_lr = False

    def __call__(self, validation_acc):
        """
        Evaluate current validation accuracy and determine if learning rate should be dropped.
        
        Args:
            validation_acc (float): Current epoch's validation accuracy
            
        Returns:
            bool: True if learning rate should be dropped, False otherwise
        """
        delta = 1 + self.delta_factor
        
        if validation_acc >= delta * self.max_validation_accuracy :
            self.max_validation_accuracy = validation_acc
            self.counter = 0  
            self.drop_lr = False
        elif self.counter < self.tolerance:
            self.counter += 1
            self.drop_lr = False
        else:
            self.counter = 0
            self.drop_lr = True

        return self.drop_lr