Method backward

code/chapter9.py:678–726

池化层的反向传播，原理见上文。 参数说明： dLdy：关于损失的梯度，为 (n_samples, out_rows, out_cols, out_ch) retain_grads：是否计算中间变量的参数梯度，bool型 输出说明： dXs：即dX，当前卷积层对输入关于损失的梯度，为 (n_samples, in_rows, in_cols, in_ch)

(self, dLdy, retain_grads=True)

        return Y

    def backward(self, dLdy, retain_grads=True):
        """
        池化层的反向传播，原理见上文。

        参数说明：
        dLdy：关于损失的梯度，为 (n_samples, out_rows, out_cols, out_ch) 
        retain_grads：是否计算中间变量的参数梯度，bool型

        输出说明：
        dXs：即dX，当前卷积层对输入关于损失的梯度，为 (n_samples, in_rows, in_cols, in_ch)
        """
        if not isinstance(dLdy, list):
            dLdy = [dLdy]

        Xs = self.X
        out_rows = self.derived_variables["out_rows"]
        out_cols = self.derived_variables["out_cols"]

        (fr, fc), s, p = self.kernel_shape, self.stride, self.pad

        dXs = []
        for X, dy, out_row, out_col in zip(Xs, dLdy, out_rows, out_cols):
            n_samp, in_rows, in_cols, nc_in = X.shape
            X_pad, (pr1, pr2, pc1, pc2) = pad2D(X, p, self.kernel_shape, s)

            dX = np.zeros_like(X_pad)
            for m in range(n_samp):
                for i in range(out_row):
                    for j in range(out_col):
                        for c in range(self.out_ch):
                            i0, i1 = i * s, (i * s) + fr
                            j0, j1 = j * s, (j * s) + fc

                            if self.mode == "max":
                                xi = X[m, i0:i1, j0:j1, c]
                                mask = np.zeros_like(xi).astype(bool)
                                x, y = np.argwhere(xi == np.max(xi))[0]
                                mask[x, y] = True
                                dX[m, i0:i1, j0:j1, c] += mask * dy[m, i, j, c]
                                
                            elif self.mode == "average":
                                frame = np.ones((fr, fc)) * dy[m, i, j, c]
                                dX[m, i0:i1, j0:j1, c] += frame / np.prod((fr, fc))

            pr2 = None if pr2 == 0 else -pr2
            pc2 = None if pc2 == 0 else -pc2
            dXs.append(dX[:, pr1:pr2, pc1:pc2, :])
            
        return dXs[0] if len(Xs) == 1 else dXs

    @property
    def hyperparams(self):