Calculate DZ Wrapper Method

Functions

calculate_dZ_wrapper(model, i, Y, loss)

Wrapper function to calculate dZ for the supported layer and loss combinations.

Parameters:

Name	Type	Description	Default
model	Sequential	The neural network model.	required
i	int	Index of the current layer.	required
Y	ndarray	True labels.	required
loss	str	Loss function used.	required

Returns:

Type	Description
	numpy.ndarray: Gradient of Z for the specified layer.

Example

dZ = calculate_dZ_wrapper(model, 1, Y, 'categorical_crossentropy')
print(dZ.shape)

Source code in microkeras/operations/backward/calculate_dZ_wrapper.py

def calculate_dZ_wrapper(model, i, Y, loss):
    """
    Wrapper function to calculate dZ for the supported layer and loss combinations.

    Args:
        model (Sequential): The neural network model.
        i (int): Index of the current layer.
        Y (numpy.ndarray): True labels.
        loss (str): Loss function used.

    Returns:
        numpy.ndarray: Gradient of Z for the specified layer.

    Example:
        ```python
        dZ = calculate_dZ_wrapper(model, 1, Y, 'categorical_crossentropy')
        print(dZ.shape)
        ```
    """
    current_layer = model.layers[i]

    if (i == len(model.layers) - 1):
        if (current_layer.activation == 'softmax' and
            loss == 'categorical_crossentropy'):
            return calculate_dZ_softmax_categorical_crossentropy(current_layer.A, Y)
        elif (current_layer.activation == 'relu' and
              loss == 'mean_squared_error'):
            return calculate_dZ_relu_mean_squared_error(current_layer.A, Y, current_layer.Z)
        elif (current_layer.activation == 'linear' and
              loss == 'mean_squared_error'):
            return calculate_dZ_linear_mean_squared_error(current_layer.A, Y, current_layer.Z)

    # General case
    next_layer = model.layers[i + 1]
    return calculate_dZ(next_layer.W,
                        next_layer.dZ,
                        current_layer.Z,
                        current_layer.activation)