[Python知识库]point cloud saliency map在rscnn上的实现

pointcloud saliency map的源代码是使用tensorflow实现的，将其移植到pytorch(0.4.1)版本的RSCNN。

环境：unbuntu16.04+cuda9.0+cudnn7.6+pytorch0.4.1+torchvizion0.2.2

实现代码：

// input :点云  分类模型  标签
//output: 丢弃高分点后的点云
def drop_points(pointclouds_pl, model, target):
    pointclouds_pl_adv_list = []
    k_list = [5, 10, 15, 20, 25, 30, 35, 40, 45, 50, 55, 60] # 分k_list次 丢弃 K_list * 5个点
    for k in k_list:

        pointclouds_pl_adv = pointclouds_pl.clone()

        for i in range(k):

            pre = model(pointclouds_pl_adv)
            loss = criterion(pre, target)
            // 输出（loss）对输入(point cloud)求倒数
            grad = torch.autograd.grad(outputs=loss, inputs=pointclouds_pl_adv,  only_inputs=True, retain_graph=False)[0]  # grad_outputs=None,

            grad = grad.cpu()
            grad = grad.detach().numpy()
            pointclouds_pl_adv = pointclouds_pl_adv.cpu()
            pointclouds_pl_adv = pointclouds_pl_adv.detach().numpy()

            sphere_core = np.median(pointclouds_pl_adv, axis=1, keepdims=True)

            sphere_r = np.sqrt(np.sum(np.square(pointclouds_pl_adv - sphere_core), axis=2))  ## BxN

            sphere_axis = pointclouds_pl_adv - sphere_core  ## BxNx3

            if 0:
                sphere_map = np.multiply(np.sum(np.multiply(grad, sphere_axis), axis=2),
                                         np.power(sphere_r, FLAGS.power))
                # sphere_map = np.multiply(np.sum(np.multiply(grad, sphere_axis), axis=2), 1/sphere_r)
                # sphere_map = np.sum(np.abs(grad), axis=2)
            else:
                sphere_map = -np.multiply(np.sum(np.multiply(grad, sphere_axis), axis=2), np.power(sphere_r, 1))
                # sphere_map = -np.multiply(np.sum(np.multiply(grad, sphere_axis), axis=2), 1/sphere_r)
                #sphere_map = np.sum(np.abs(grad), axis=2)

            drop_indice = np.argpartition(sphere_map, kth=sphere_map.shape[1] - 5, axis=1)[:, -5:]

            tmp = np.zeros((pointclouds_pl_adv.shape[0], pointclouds_pl_adv.shape[1] - 5, 3), dtype=float)
            for j in range(pointclouds_pl.shape[0]):
                tmp[j] = np.delete(pointclouds_pl_adv[j], drop_indice[j], axis=0)  # along N points to delete


            pointclouds_pl_adv = tmp.copy()

            pointclouds_pl_adv = torch.from_numpy(pointclouds_pl_adv)
            pointclouds_pl_adv = pointclouds_pl_adv.type(torch.float32)
            pointclouds_pl_adv = pointclouds_pl_adv.cuda()
            pointclouds_pl_adv.requires_grad_()

        pointclouds_pl_adv_list.append(pointclouds_pl_adv)

    return pointclouds_pl_adv_list

在运行grad = torch.autograd.grad(outputs=loss, inputs=pointclouds_pl_adv, only_inputs=True, retain_graph=False)[0] # grad_outputs=None, 时候会报错，错误 大概是 BallQueryBackward 返回了错误数量的gradient (expecte 5, got 4)

这时将文件 pointnet2_utils.py中的

    def backward(ctx, a=None):
        #print('ball backward')

        return None, None, None, None

改为（多加一个None），即可正常运行

    def backward(ctx, a=None):
        #print('ball backward')

        return None, None, None, None, None

超强干货来袭

云风专访：近40年码龄，通宵达旦的技术人生