【python實現卷積神經網絡】padding2D層實現

代碼來源：https://github.com/eriklindernoren/ML-From-Scratch

卷積神經網絡中卷積層Conv2D（帶stride、padding）的具體實現：https://www.cnblogs.com/xiximayou/p/12706576.html

激活函數的實現（sigmoid、softmax、tanh、relu、leakyrelu、elu、selu、softplus）：https://www.cnblogs.com/xiximayou/p/12713081.html

損失函數定義（均方誤差、交叉熵損失）：https://www.cnblogs.com/xiximayou/p/12713198.html

優化器的實現（SGD、Nesterov、Adagrad、Adadelta、RMSprop、Adam）：https://www.cnblogs.com/xiximayou/p/12713594.html

卷積層反向傳播過程：https://www.cnblogs.com/xiximayou/p/12713930.html

全連接層實現：https://www.cnblogs.com/xiximayou/p/12720017.html

批量歸一化層實現：https://www.cnblogs.com/xiximayou/p/12720211.html

池化層實現：https://www.cnblogs.com/xiximayou/p/12720324.html

 

class ConstantPadding2D(Layer):
    """Adds rows and columns of constant values to the input.
    Expects the input to be of shape (batch_size, channels, height, width)
    Parameters:
    -----------
    padding: tuple
        The amount of padding along the height and width dimension of the input.
        If (pad_h, pad_w) the same symmetric padding is applied along height and width dimension.
        If ((pad_h0, pad_h1), (pad_w0, pad_w1)) the specified padding is added to beginning and end of
        the height and width dimension.
    padding_value: int or tuple
        The value the is added as padding.
    """
    def __init__(self, padding, padding_value=0):
        self.padding = padding
        self.trainable = True
        if not isinstance(padding[0], tuple):
            self.padding = ((padding[0], padding[0]), padding[1])
        if not isinstance(padding[1], tuple):
            self.padding = (self.padding[0], (padding[1], padding[1]))
        self.padding_value = padding_value

    def forward_pass(self, X, training=True):
        output = np.pad(X,
            pad_width=((0,0), (0,0), self.padding[0], self.padding[1]),
            mode="constant",
            constant_values=self.padding_value)
        return output

    def backward_pass(self, accum_grad):
        pad_top, pad_left = self.padding[0][0], self.padding[1][0]
        height, width = self.input_shape[1], self.input_shape[2]
        accum_grad = accum_grad[:, :, pad_top:pad_top+height, pad_left:pad_left+width]
        return accum_grad

    def output_shape(self):
        new_height = self.input_shape[1] + np.sum(self.padding[0])
        new_width = self.input_shape[2] + np.sum(self.padding[1])
        return (self.input_shape[0], new_height, new_width)


class ZeroPadding2D(ConstantPadding2D):
    """Adds rows and columns of zero values to the input.
    Expects the input to be of shape (batch_size, channels, height, width)
    Parameters:
    -----------
    padding: tuple
        The amount of padding along the height and width dimension of the input.
        If (pad_h, pad_w) the same symmetric padding is applied along height and width dimension.
        If ((pad_h0, pad_h1), (pad_w0, pad_w1)) the specified padding is added to beginning and end of
        the height and width dimension.
    """
    def __init__(self, padding):
        self.padding = padding
        if isinstance(padding[0], int):
            self.padding = ((padding[0], padding[0]), padding[1])
        if isinstance(padding[1], int):
            self.padding = (self.padding[0], (padding[1], padding[1]))
        self.padding_value = 0

需要注意的是輸入的維度是：[batchsize,channel,height,width]，因此在進行padding的時候是在最后兩個維度上進行操作的。

假設輸入的圖像維度為[1,3,32,32]，輸入的padding=((1,1),(1,1))，accm_grad是后一層傳到該層的梯度，那么padding2D的反向傳播的梯度accm_grad=accm_grad[:, :, 1:33, 1:33]。