classification

以分10类，gpu版本为例。
先将pytorch训练出的pth权重文件转为onnx文件：

import torch
import pointnet_cls


point_num = 1024
class_num = 10
normal_channel = False

model = pointnet_cls.get_model(class_num, normal_channel)
model = model.cuda() #cpu版本需注释此句
model.eval()
checkpoint = torch.load('./cls.pth')
model.load_state_dict(checkpoint['model_state_dict'])

x = (torch.rand(1, 6, point_num) if normal_channel else torch.rand(1, 3, point_num))
x = x.cuda() #cpu版本需注释此句

export_onnx_file = "./cls.onnx"
torch.onnx.export(model,
                    x,
                    export_onnx_file,
                    opset_version = 11)

python推理：

import numpy as np
import onnxruntime


point_num = 1024

 
def pc_normalize(pc):
    centroid = np.mean(pc, axis=0)
    pc = pc - centroid
    m = np.max(np.sqrt(np.sum(pc**2, axis=1)))
    pc = pc / m
    return pc


if __name__ == '__main__':
    file = './bed_0610.txt'
    data = np.loadtxt(file, delimiter=',').astype(np.float32)
    point_set = data[:, 0:3]
    point_set = point_set[0:point_num, :]     
    point_set[:, 0:3] = pc_normalize(point_set[:, 0:3])
    points = np.reshape(point_set, ((1, point_num, 3)))
    points = points.swapaxes(2, 1)

    onnx_session = onnxruntime.InferenceSession("cls.onnx", providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])

    input_name = []
    for node in onnx_session.get_inputs():
        input_name.append(node.name)

    output_name = []
    for node in onnx_session.get_outputs():
        output_name.append(node.name)

    inputs = {}
    for name in input_name:
        inputs[name] = points

    outputs = onnx_session.run(None, inputs)[0]
    print(np.argmax(outputs))

C++推理：

#include <iostream>
#include <vector>
#include <fstream>
#include <onnxruntime_cxx_api.h>


const int point_num = 1024;
const int class_num = 10;


void pc_normalize(std::vector<float>& points)
{
	float mean_x = 0, mean_y = 0, mean_z = 0;
	for (size_t i = 0; i < point_num; ++i)
	{
		mean_x += points[3 * i];
		mean_y += points[3 * i + 1];
		mean_z += points[3 * i + 2];
	}
	mean_x /= point_num;
	mean_y /= point_num;
	mean_z /= point_num;

	for (size_t i = 0; i < point_num; ++i)
	{
		points[3 * i] -= mean_x;
		points[3 * i + 1] -= mean_y;
		points[3 * i + 2] -= mean_z;
	}

	float m = 0;
	for (size_t i = 0; i < point_num; ++i)
	{
		if (sqrt(pow(points[3 * i], 2) + pow(points[3 * i + 1], 2) + pow(points[3 * i + 2], 2)) > m)
			m = sqrt(pow(points[3 * i], 2) + pow(points[3 * i + 1], 2) + pow(points[3 * i + 2], 2));
	}

	for (size_t i = 0; i < point_num; ++i)
	{
		points[3 * i] /= m;
		points[3 * i + 1] /= m;
		points[3 * i + 2] /= m;
	}
}


void classfier(std::vector<float> & points)
{
	Ort::Env env(ORT_LOGGING_LEVEL_WARNING, "cls");
	Ort::SessionOptions session_options;
	session_options.SetIntraOpNumThreads(1);
	session_options.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_EXTENDED);

	OrtCUDAProviderOptions cuda_option;
	cuda_option.device_id = 0;
	cuda_option.arena_extend_strategy = 0;
	cuda_option.cudnn_conv_algo_search = OrtCudnnConvAlgoSearchExhaustive;
	cuda_option.gpu_mem_limit = SIZE_MAX;
	cuda_option.do_copy_in_default_stream = 1;
	session_options.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_ALL);
	session_options.AppendExecutionProvider_CUDA(cuda_option);

	const wchar_t* model_path = L"cls.onnx";
	Ort::Session session(env, model_path, session_options);
	Ort::AllocatorWithDefaultOptions allocator;

	std::vector<const char*>  input_node_names;
	for (size_t i = 0; i < session.GetInputCount(); i++)
	{
		input_node_names.push_back(session.GetInputName(i, allocator));
	}

	std::vector<const char*> output_node_names;
	for (size_t i = 0; i < session.GetOutputCount(); i++)
	{
		output_node_names.push_back(session.GetOutputName(i, allocator));
	}

	const size_t input_tensor_size = 1 * 3 * point_num ;
	std::vector<float> input_tensor_values(input_tensor_size);
	for (size_t i = 0; i < 3; i++)
	{
		for (size_t j = 0; j < point_num; j++)
		{
			input_tensor_values[point_num * i + j] = points[3 * j + i];
		}
	}

	std::vector<int64_t> input_node_dims = { 1, 3, point_num };
	auto memory_info = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
	Ort::Value input_tensor = Ort::Value::CreateTensor<float>(memory_info, input_tensor_values.data(), input_tensor_size, input_node_dims.data(), input_node_dims.size());

	std::vector<Ort::Value> inputs;
	inputs.push_back(std::move(input_tensor));

	std::vector<Ort::Value> outputs = session.Run(Ort::RunOptions{ nullptr }, input_node_names.data(), inputs.data(), input_node_names.size(), output_node_names.data(), output_node_names.size());

	const float* rawOutput = outputs[0].GetTensorData<float>();
	std::vector<int64_t> outputShape = outputs[0].GetTensorTypeAndShapeInfo().GetShape();
	size_t count = outputs[0].GetTensorTypeAndShapeInfo().GetElementCount();
	std::vector<float> pred(rawOutput, rawOutput + count);

	int predict_label = std::max_element(pred.begin(), pred.end()) - pred.begin();
	std::cout << predict_label << std::endl;
}


int main()
{
	std::vector<float> points;
	float x, y, z, nx, ny, nz;
	char ch;
	std::ifstream infile("bed_0610.txt");
	for (size_t i = 0; i < point_num; i++)
	{
		infile >> x >> ch >> y >> ch >> z >> ch >> nx >> ch >> ny >> ch >> nz;
		points.push_back(x);
		points.push_back(y);
		points.push_back(z);
	}
	infile.close();

	pc_normalize(points);

	classfier(points);

	return 0;
}

part segmentation

以分16类50部分，gpu版本为例。
先将pytorch训练出的pth权重文件转为onnx文件：

import torch
import pointnet_part_seg


point_num = 2048
class_num = 16
part_num = 50
normal_channel = False

def to_categorical(y, class_num):
    """ 1-hot encodes a tensor """
    new_y = torch.eye(class_num)[y.cpu().data.numpy(),]
    if (y.is_cuda):
        return new_y.cuda()
    return new_y

model = pointnet_part_seg.get_model(part_num, normal_channel)
model = model.cuda() #cpu版本需注释此句
model.eval()
checkpoint = torch.load('./part_seg.pth')
model.load_state_dict(checkpoint['model_state_dict'])

x = (torch.rand(1, 6, point_num) if normal_channel else torch.rand(1, 3, point_num))
x = x.cuda() #cpu版本需注释此句
label = torch.randint(0, 1, (1, 1))
label = label.cuda() #cpu版本需注释此句

export_onnx_file = "./part_seg.onnx"			
torch.onnx.export(model,
                    (x, to_categorical(label, class_num)),
                    export_onnx_file,
                    opset_version = 11)

python推理：

import numpy as np
import onnxruntime


point_num = 2048
class_num = 16

 
def to_categorical(y, class_num):
    """ 1-hot encodes a tensor """
    new_y = np.eye(class_num)[y,]
    return new_y.astype(np.float32)


def pc_normalize(pc):
    centroid = np.mean(pc, axis=0)
    pc = pc - centroid
    m = np.max(np.sqrt(np.sum(pc ** 2, axis=1)))
    pc = pc / m
    return pc


if __name__ == '__main__':
    data = np.loadtxt('85a15c26a6e9921ae008cc4902bfe3cd.txt').astype(np.float32)
    point_set = data[:, 0:3]
    point_set[:, 0:3] = pc_normalize(point_set[:, 0:3])

    choice = np.random.choice(point_set.shape[0], point_num, replace=True)
    point_set = point_set[choice, :][:, 0:3]
    pts = point_set

    points = np.reshape(point_set, ((1, point_num, 3)))
    points = points.swapaxes(2, 1)
    label = np.array([[0]], dtype=np.int32)

    onnx_session = onnxruntime.InferenceSession("part_seg.onnx", providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])

    input_name = []
    for node in onnx_session.get_inputs():
        input_name.append(node.name)

    output_name = []
    for node in onnx_session.get_outputs():
        output_name.append(node.name)

    inputs = {}
    inputs[input_name[0]] = points
    inputs[input_name[1]] = to_categorical(label, class_num)

    outputs = onnx_session.run(None, inputs)[0]

    cur_pred_val_logits = outputs
    cur_pred_val = np.zeros((1, point_num)).astype(np.int32)
    
    logits = cur_pred_val_logits[0, :, :]
    cur_pred_val[0, :] = np.argmax(logits, 1)

    pts = np.append(pts.reshape(point_num, 3), cur_pred_val[0, :].reshape(point_num, 1), 1)
    np.savetxt('pred.txt', pts, fmt='%.06f')       

C++推理：

#include <iostream>
#include <vector>
#include <fstream>
#include <ctime>
#include <onnxruntime_cxx_api.h>


const int point_num = 2048;
const int class_num = 16;
const int parts_num = 50;


void pc_normalize(std::vector<float>& points)
{
	float mean_x = 0, mean_y = 0, mean_z = 0;
	for (size_t i = 0; i < point_num; ++i)
	{
		mean_x += points[3 * i];
		mean_y += points[3 * i + 1];
		mean_z += points[3 * i + 2];
	}
	mean_x /= point_num;
	mean_y /= point_num;
	mean_z /= point_num;

	for (size_t i = 0; i < point_num; ++i)
	{
		points[3 * i] -= mean_x;
		points[3 * i + 1] -= mean_y;
		points[3 * i + 2] -= mean_z;
	}

	float m = 0;
	for (size_t i = 0; i < point_num; ++i)
	{
		if (sqrt(pow(points[3 * i], 2) + pow(points[3 * i + 1], 2) + pow(points[3 * i + 2], 2)) > m)
			m = sqrt(pow(points[3 * i], 2) + pow(points[3 * i + 1], 2) + pow(points[3 * i + 2], 2));
	}

	for (size_t i = 0; i < point_num; ++i)
	{
		points[3 * i] /= m;
		points[3 * i + 1] /= m;
		points[3 * i + 2] /= m;
	}
}


void resample(std::vector<float>& points)
{
	srand((int)time(0));
	std::vector<int> choice(point_num);
	for (size_t i = 0; i < point_num; i++)
	{
		choice[i] = rand() % (points.size() / 3);
	}

	std::vector<float> temp_points(3 * point_num);
	for (size_t i = 0; i < point_num; i++)
	{
		temp_points[3 * i] = points[3 * choice[i]];
		temp_points[3 * i + 1] = points[3 * choice[i] + 1];
		temp_points[3 * i + 2] = points[3 * choice[i] + 2];
	}
	points = temp_points;
}


std::vector<int> classfier(std::vector<float> & points, std::vector<float> & labels)
{
	Ort::Env env(ORT_LOGGING_LEVEL_WARNING, "part_seg");
	Ort::SessionOptions session_options;
	session_options.SetIntraOpNumThreads(1);
	session_options.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_EXTENDED);

	OrtCUDAProviderOptions cuda_option;
	cuda_option.device_id = 0;
	cuda_option.arena_extend_strategy = 0;
	cuda_option.cudnn_conv_algo_search = OrtCudnnConvAlgoSearchExhaustive;
	cuda_option.gpu_mem_limit = SIZE_MAX;
	cuda_option.do_copy_in_default_stream = 1;
	session_options.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_ALL);
	session_options.AppendExecutionProvider_CUDA(cuda_option);

	const wchar_t* model_path = L"part_seg.onnx";
	Ort::Session session(env, model_path, session_options);
	Ort::AllocatorWithDefaultOptions allocator;

	std::vector<const char*>  input_node_names;
	for (size_t i = 0; i < session.GetInputCount(); i++)
	{
		input_node_names.push_back(session.GetInputName(i, allocator));
	}

	std::vector<const char*> output_node_names;
	for (size_t i = 0; i < session.GetOutputCount(); i++)
	{
		output_node_names.push_back(session.GetOutputName(i, allocator));
	}

	const size_t input_tensor_size0 = 1 * 3 * point_num;
	std::vector<float> input_tensor_values0(input_tensor_size0);
	for (size_t i = 0; i < 3; i++)
	{
		for (size_t j = 0; j < point_num; j++)
		{
			input_tensor_values0[point_num * i + j] = points[3 * j + i];
		}
	}
	std::vector<int64_t> input_node_dims0 = { 1, 3, point_num };
	auto memory_info0 = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
	Ort::Value input_tensor0 = Ort::Value::CreateTensor<float>(memory_info0, input_tensor_values0.data(), input_tensor_size0, input_node_dims0.data(), input_node_dims0.size());

	const size_t input_tensor_size1 = 1 * 1 * class_num;
	std::vector<float> input_tensor_values1(input_tensor_size0);
	for (size_t i = 0; i < class_num; i++)
	{
		input_tensor_values1[i] = labels[i];
	}
	std::vector<int64_t> input_node_dims1 = { 1, 1, class_num };
	auto memory_info1 = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
	Ort::Value input_tensor1 = Ort::Value::CreateTensor<float>(memory_info1, input_tensor_values1.data(), input_tensor_size1, input_node_dims1.data(), input_node_dims1.size());

	std::vector<Ort::Value> inputs;
	inputs.push_back(std::move(input_tensor0));
	inputs.push_back(std::move(input_tensor1));

	std::vector<Ort::Value> outputs = session.Run(Ort::RunOptions{ nullptr }, input_node_names.data(), inputs.data(), input_node_names.size(), output_node_names.data(), output_node_names.size());

	const float* rawOutput = outputs[0].GetTensorData<float>();
	std::vector<int64_t> outputShape = outputs[0].GetTensorTypeAndShapeInfo().GetShape();
	size_t count = outputs[0].GetTensorTypeAndShapeInfo().GetElementCount();
	std::vector<float> pred(rawOutput, rawOutput + count);

	std::vector<std::vector<float>> preds(point_num, std::vector<float>(parts_num, 0));

	for (size_t i = 0; i < point_num; i++)
	{
		for (size_t j = 0; j < parts_num; j++)
		{
			preds[i][j] = pred[i * parts_num + j];
		}
	}

	std::vector<int> max_index(point_num, 0);

	for (size_t i = 0; i < point_num; i++)
	{
		max_index[i]= std::max_element(preds[i].begin(), preds[i].end()) - preds[i].begin();
	}
	return max_index;
}


int main()
{
	std::vector<float> points, labels;
	float x, y, z, nx, ny, nz, label;
	std::ifstream infile("85a15c26a6e9921ae008cc4902bfe3cd.txt");
	while (infile >> x >> y >> z >> nx >> ny >> nz >> label)
	{
		points.push_back(x);
		points.push_back(y);
		points.push_back(z);
	}
	for (size_t i = 0; i < class_num; i++)
	{
		labels.push_back(0.0);
	}
	labels[0] = 1.0;
	infile.close();

	pc_normalize(points);

	resample(points);

	std::vector<int> result = classfier(points, labels);

	std::fstream outfile("pred.txt", 'w');
	for (size_t i = 0; i < point_num; i++)
	{
		outfile << points[3 * i] << " " << points[3 * i + 1] << " " << points[3 * i + 2] << " " << result[i]<< std::endl;
	}
	outfile.close();

	return 0;
}

sematic segmentation

以分13类，gpu版本为例。
先将pytorch训练出的pth权重文件转为onnx文件：

import torch
import pointnet_sem_seg


point_num = 4096
class_num = 13

model = pointnet_sem_seg.get_model(class_num)
model = model.cuda() #cpu版本需注释此句
model.eval()
checkpoint = torch.load('sem_seg.pth')
model.load_state_dict(checkpoint['model_state_dict'])

x = torch.rand(1, 9, point_num)
x = x.cuda() #cpu版本需注释此句

export_onnx_file = "./sem_seg.onnx"
torch.onnx.export(model,
                    x,
                    export_onnx_file,
                    opset_version = 11)

python推理代码：

import numpy as np
import onnxruntime


point_num = 4096
class_num = 13
stride = 0.5
block_size = 1.0


if __name__ == '__main__':
    data = np.load('Area_1_conferenceRoom_1.npy')
    points = data[:,:6]
    coord_min, coord_max = np.amin(points, axis=0)[:3], np.amax(points, axis=0)[:3]
    grid_x = int(np.ceil(float(coord_max[0] - coord_min[0] - block_size) / stride) + 1)
    grid_y = int(np.ceil(float(coord_max[1] - coord_min[1] - block_size) / stride) + 1)
    data_room, index_room = np.array([]), np.array([])
    for index_y in range(0, grid_y):
        for index_x in range(0, grid_x):
            s_x = coord_min[0] + index_x * stride
            e_x = min(s_x + block_size, coord_max[0])
            s_x = e_x - block_size
            s_y = coord_min[1] + index_y * stride
            e_y = min(s_y + block_size, coord_max[1])
            s_y = e_y - block_size
            point_idxs = np.where((points[:, 0] >= s_x) & (points[:, 0] <= e_x) & (points[:, 1] >= s_y) & (points[:, 1] <= e_y))[0]
            if point_idxs.size == 0:
                continue
            num_batch = int(np.ceil(point_idxs.size / point_num))
            point_size = int(num_batch * point_num)
            replace = False if (point_size - point_idxs.size <= point_idxs.size) else True
            point_idxs_repeat = np.random.choice(point_idxs, point_size - point_idxs.size, replace=replace)
            point_idxs = np.concatenate((point_idxs, point_idxs_repeat))
            np.random.shuffle(point_idxs)
            data_batch = points[point_idxs, :]
            normlized_xyz = np.zeros((point_size, 3))
            normlized_xyz[:, 0] = data_batch[:, 0] / coord_max[0]
            normlized_xyz[:, 1] = data_batch[:, 1] / coord_max[1]
            normlized_xyz[:, 2] = data_batch[:, 2] / coord_max[2]
            data_batch[:, 0] = data_batch[:, 0] - (s_x + block_size / 2.0)
            data_batch[:, 1] = data_batch[:, 1] - (s_y + block_size / 2.0)
            data_batch[:, 3:6] /= 255.0
            data_batch = np.concatenate((data_batch, normlized_xyz), axis=1)
            data_room = np.vstack([data_room, data_batch]) if data_room.size else data_batch
            index_room = np.hstack([index_room, point_idxs]) if index_room.size else point_idxs
    data_room = data_room.reshape((-1, point_num, data_room.shape[1]))
    index_room = index_room.reshape((-1, point_num))

    onnx_session = onnxruntime.InferenceSession("sem_seg.onnx", providers=['CUDAExecutionProvider', 'CPUExecutionProvider'])

    input_name = []
    for node in onnx_session.get_inputs():
        input_name.append(node.name)

    output_name = []
    for node in onnx_session.get_outputs():
        output_name.append(node.name)

    vote_label_pool = np.zeros((points.shape[0], class_num))
    num_blocks = data_room.shape[0]
    batch_data = np.zeros((1, point_num, 9))
    batch_point_index = np.zeros((1, point_num))

    for sbatch in range(num_blocks):
        start_idx = sbatch
        end_idx = min(sbatch + 1, num_blocks)
        real_batch_size = end_idx - start_idx
        batch_data[0:real_batch_size, ...] = data_room[start_idx:end_idx, ...]
        batch_point_index[0:real_batch_size, ...] = index_room[start_idx:end_idx, ...]

        inputs = {}
        for name in input_name:
            inputs[name] = batch_data.swapaxes(2, 1).astype(np.float32)

        outputs = onnx_session.run(None, inputs)[0]

        batch_pred_label = np.argmax(outputs, 2)

        point_idx = batch_point_index[0:real_batch_size, ...]
        pred_label = batch_pred_label[0:real_batch_size, ...]
        for b in range(pred_label.shape[0]):
            for n in range(pred_label.shape[1]):
                vote_label_pool[int(point_idx[b, n]), int(pred_label[b, n])] += 1

    pred = np.argmax(vote_label_pool, 1)
    fout = open('pred.txt', 'w')
    for i in range(points.shape[0]):
        fout.write('%f %f %f %d\n' % (points[i, 0], points[i, 1], points[i, 2], pred[i]))
    fout.close()

C++推理：

#include <iostream>
#include <fstream>
#include <vector>
#include <algorithm>
#include <ctime>
#include <random>
#include <onnxruntime_cxx_api.h>


const int point_num = 4096;
const int class_num = 13;


struct point
{
	float m_x, m_y, m_z, m_r, m_g, m_b, m_normal_x, m_normal_y, m_normal_z;
	point() :
		m_x(0), m_y(0), m_z(0), m_r(0), m_g(0), m_b(0), m_normal_x(0), m_normal_y(0), m_normal_z(0) {}
	point(float x, float y, float z, float r, float g, float b) :
		m_x(x), m_y(y), m_z(z), m_r(r), m_g(g), m_b(b), m_normal_x(0), m_normal_y(0), m_normal_z(0) {}
	point(float x, float y, float z, float r, float g, float b, float normal_x, float normal_y, float normal_z) :
		m_x(x), m_y(y), m_z(z), m_r(r), m_g(g), m_b(b), m_normal_x(normal_x), m_normal_y(normal_y), m_normal_z(normal_z) {}
};


int main()
{
	float x, y, z, r, g, b, l;
	std::vector<point> pts;
	std::vector<float> points_x, points_y, points_z;
	int points_num = 0;
	std::ifstream infile("Area_1_conferenceRoom_1.txt");
	while (infile >> x >> y >> z >> r >> g >> b >> l)
	{
		point pt(x, y, z, r, g, b);
		pts.push_back(pt);
		points_x.push_back(x);
		points_y.push_back(y);
		points_z.push_back(z);
		points_num++;
	}

	float x_min = *std::min_element(points_x.begin(), points_x.end());
	float y_min = *std::min_element(points_y.begin(), points_y.end());
	float z_min = *std::min_element(points_z.begin(), points_z.end());
	float x_max = *std::max_element(points_x.begin(), points_x.end());
	float y_max = *std::max_element(points_y.begin(), points_y.end());
	float z_max = *std::max_element(points_z.begin(), points_z.end());

	float stride = 0.5;
	float block_size = 1.0;
	srand((int)time(0));

	int grid_x = ceil((x_max - x_min - block_size) / stride) + 1;
	int grid_y = ceil((y_max - y_min - block_size) / stride) + 1;

	std::vector<point> data_room;
	std::vector<int> index_room;
	for (size_t index_y = 0; index_y < grid_y; index_y++)
	{
		for (size_t index_x = 0; index_x < grid_x; index_x++)
		{
			float s_x = x_min + index_x * stride;
			float e_x = std::min(s_x + block_size, x_max);
			s_x = e_x - block_size;
			float s_y = y_min + index_y * stride;
			float e_y = std::min(s_y + block_size, y_max);
			s_y = e_y - block_size;

			std::vector<int> point_idxs;
			for (size_t i = 0; i < points_num; i++)
			{
				if (points_x[i] >= s_x && points_x[i] <= e_x && points_y[i] >= s_y && points_y[i] <= e_y)
					point_idxs.push_back(i);
			}
			if (point_idxs.size() == 0)
				continue;

			int num_batch = ceil(point_idxs.size() * 1.0 / point_num);
			int point_size = num_batch * point_num;
			bool replace = (point_size - point_idxs.size() <= point_idxs.size() ? false : true);

			std::vector<int> point_idxs_repeat;
			if (replace)
			{
				for (size_t i = 0; i < point_size - point_idxs.size(); i++)
				{
					int id = rand() % point_idxs.size();
					point_idxs_repeat.push_back(point_idxs[id]);
				}
			}
			else
			{
				std::vector<bool> flags(pts.size(), false);
				for (size_t i = 0; i < point_size - point_idxs.size(); i++)
				{
					int id = rand() % point_idxs.size();
					while (true)
					{
						if (flags[id] == false)
						{
							flags[id] = true;
							break;
						}
						id = rand() % point_idxs.size();
					}
					point_idxs_repeat.push_back(point_idxs[id]);
				}
			}
			point_idxs.insert(point_idxs.end(), point_idxs_repeat.begin(), point_idxs_repeat.end());

			std::random_device rd;
			std::mt19937 g(rd());	// 随机数引擎:基于梅森缠绕器算法的随机数生成器
			std::shuffle(point_idxs.begin(), point_idxs.end(), g);	// 打乱顺序，重新排序（随机序列）

			std::vector<point> data_batch;
			for (size_t i = 0; i < point_idxs.size(); i++)
			{
				data_batch.push_back(pts[point_idxs[i]]);
			}

			for (size_t i = 0; i < point_size; i++)
			{
				data_batch[i].m_normal_x = data_batch[i].m_x / x_max;
				data_batch[i].m_normal_y = data_batch[i].m_y / y_max;
				data_batch[i].m_normal_z = data_batch[i].m_z / z_max;
				data_batch[i].m_x -= (s_x + block_size / 2.0);
				data_batch[i].m_y -= (s_y + block_size / 2.0);
				data_batch[i].m_r /= 255.0;
				data_batch[i].m_g /= 255.0;
				data_batch[i].m_b /= 255.0;
				data_room.push_back(data_batch[i]);
				index_room.push_back(point_idxs[i]);
			}
		}
	}

	int n = point_num, m = index_room.size() / n;
	std::vector<std::vector<point>> data_rooms(m, std::vector<point>(n, point()));
	std::vector<std::vector<int>> index_rooms(m, std::vector<int>(n, 0));
	for (size_t i = 0; i < m; i++)
	{
		for (size_t j = 0; j < n; j++)
		{
			data_rooms[i][j] = data_room[i * n + j];
			index_rooms[i][j] = index_room[i * n + j];
		}
	}

	std::vector<std::vector<int>> vote_label_pool(points_num, std::vector<int>(class_num, 0));
	int num_blocks = data_rooms.size();

	Ort::Env env(ORT_LOGGING_LEVEL_WARNING, "sem_seg");
	Ort::SessionOptions session_options;
	session_options.SetIntraOpNumThreads(1);
	session_options.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_EXTENDED);

	OrtCUDAProviderOptions cuda_option;
	cuda_option.device_id = 0;
	cuda_option.arena_extend_strategy = 0;
	cuda_option.cudnn_conv_algo_search = OrtCudnnConvAlgoSearchExhaustive;
	cuda_option.gpu_mem_limit = SIZE_MAX;
	cuda_option.do_copy_in_default_stream = 1;
	session_options.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_ALL);
	session_options.AppendExecutionProvider_CUDA(cuda_option);

	const wchar_t* model_path = L"sem_seg.onnx";
	Ort::Session session(env, model_path, session_options);
	Ort::AllocatorWithDefaultOptions allocator;

	std::vector<const char*>  input_node_names;
	for (size_t i = 0; i < session.GetInputCount(); i++)
	{
		input_node_names.push_back(session.GetInputName(i, allocator));
	}

	std::vector<const char*> output_node_names;
	for (size_t i = 0; i < session.GetOutputCount(); i++)
	{
		output_node_names.push_back(session.GetOutputName(i, allocator));
	}

	const size_t input_tensor_size = 1 * 9 * point_num;
	std::vector<float> input_tensor_values(input_tensor_size);

	for (int sbatch = 0; sbatch < num_blocks; sbatch++)
	{
		int start_idx = sbatch;
		int end_idx = std::min(sbatch + 1, num_blocks);
		int real_batch_size = end_idx - start_idx;
		std::vector<point> batch_data = data_rooms[start_idx];
		std::vector<int> point_idx = index_rooms[start_idx];
		std::vector<float> batch(point_num * 9);
		for (size_t i = 0; i < point_num; i++)
		{
			batch[9 * i + 0] = batch_data[i].m_x;
			batch[9 * i + 1] = batch_data[i].m_y;
			batch[9 * i + 2] = batch_data[i].m_z;
			batch[9 * i + 3] = batch_data[i].m_r;
			batch[9 * i + 4] = batch_data[i].m_g;
			batch[9 * i + 5] = batch_data[i].m_b;
			batch[9 * i + 6] = batch_data[i].m_normal_x;
			batch[9 * i + 7] = batch_data[i].m_normal_y;
			batch[9 * i + 8] = batch_data[i].m_normal_z;
		}

		for (size_t i = 0; i < 9; i++)
		{
			for (size_t j = 0; j < point_num; j++)
			{
				input_tensor_values[i * point_num + j] = batch[9 * j + i];
			}
		}

		std::vector<int64_t> input_node_dims = { 1, 9, point_num };
		auto memory_info = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
		Ort::Value input_tensor = Ort::Value::CreateTensor<float>(memory_info, input_tensor_values.data(), input_tensor_size, input_node_dims.data(), input_node_dims.size());

		std::vector<Ort::Value> inputs;
		inputs.push_back(std::move(input_tensor));

		std::vector<Ort::Value> outputs = session.Run(Ort::RunOptions{ nullptr }, input_node_names.data(), inputs.data(), input_node_names.size(), output_node_names.data(), output_node_names.size());

		const float* rawOutput = outputs[0].GetTensorData<float>();
		std::vector<int64_t> outputShape = outputs[0].GetTensorTypeAndShapeInfo().GetShape();
		size_t count = outputs[0].GetTensorTypeAndShapeInfo().GetElementCount();
		std::vector<float> pred(rawOutput, rawOutput + count);

		std::vector<std::vector<float>> preds(point_num, std::vector<float>(class_num, 0));

		for (size_t i = 0; i < point_num; i++)
		{
			for (size_t j = 0; j < class_num; j++)
			{
				preds[i][j] = pred[i * class_num + j];
			}
		}

		std::vector<int> pred_label(point_num, 0);
		for (size_t i = 0; i < point_num; i++)
		{
			pred_label[i] = std::max_element(preds[i].begin(), preds[i].end()) - preds[i].begin();
			vote_label_pool[point_idx[i]][pred_label[i]] += 1;
		}
	}

	std::ofstream outfile("pred.txt");
	for (size_t i = 0; i < points_num; i++)
	{
		int max_index = std::max_element(vote_label_pool[i].begin(), vote_label_pool[i].end()) - vote_label_pool[i].begin();
		outfile << pts[i].m_x << " " << pts[i].m_y << " " << pts[i].m_z << " " << max_index << std::endl;
	}
	outfile.close();

	return 0;
}

注意，由于C++无法直接读取npy格式文件（可以依赖一些库），这里先使用python脚本将npy文件转换成txt文件。

import numpy as np

npy = np.load("Area_1_conferenceRoom_1.npy")
np.savetxt('Area_1_conferenceRoom_1.txt', npy, fmt='%0.06f')

模型的下载地址：pointnet模型权重