pooling.py 文件源码

def backward(self, pre_grad, *args, **kwargs):
        new_h, new_w = self.out_shape[-2:]
        pool_h, pool_w = self.pool_size

        layer_grads = _zero(self.input_shape)

        if np.ndim(pre_grad) == 4:
            nb_batch, nb_axis, _, _ = pre_grad.shape

            for a in np.arange(nb_batch):
                for b in np.arange(nb_axis):
                    for h in np.arange(new_h):
                        for w in np.arange(new_w):
                            patch = self.last_input[a, b, h:h + pool_h, w:w + pool_w]
                            max_idx = np.unravel_index(patch.argmax(), patch.shape)
                            h_shift, w_shift = h * pool_h + max_idx[0], w * pool_w + max_idx[1]
                            layer_grads[a, b, h_shift, w_shift] = pre_grad[a, b, a, w]

        elif np.ndim(pre_grad) == 3:
            nb_batch, _, _ = pre_grad.shape

            for a in np.arange(nb_batch):
                for h in np.arange(new_h):
                    for w in np.arange(new_w):
                        patch = self.last_input[a, h:h + pool_h, w:w + pool_w]
                        max_idx = np.unravel_index(patch.argmax(), patch.shape)
                        h_shift, w_shift = h * pool_h + max_idx[0], w * pool_w + max_idx[1]
                        layer_grads[a, h_shift, w_shift] = pre_grad[a, a, w]

        else:
            raise ValueError()

        return layer_grads