征程 6M 部署 Omnidet 感知模型

一、导出 onnx 模型

服务器端或者本地电脑 git clone github.com/valeoai/Woo...

进入 omnidet 下的 models 文件夹下面找到 onnx 文件夹，会找到 export_onnx.py 文件。根据文件顶端说明，需要去配置相关的文件。进入 data/params.yaml 中，在配置文件的最下 main 可以找到# -- ONNX MODEL EXPORT --相关配置 。choices 中可以选择你需要的任务，加入你只需要分割，就将 onnx_modelg 更改为 segmentic，假如你的任务是 detection，就将 onnx_model 更改为 detection。这里我们以分割为示例模型。将按照示例得到的训练模型路径写入到 model_path，作为导出 onnx 的输入。然后在 onnx_export_path 写 onnx 的导出路径。保存，终端运行 python . /onnx_export.py --config data/params.yaml。获取到 onnx 模型。

二、校准集准备：

import os
import sys
import argparse
import yaml
sys.path.append('./WoodScape-ICCV19/omnidet')
import cv2
import matplotlib as mpl
import matplotlib.cm as cm
import numpy as np
import onnxruntime
import torch
from PIL import Image
from matplotlib import pyplot as plt
from horizon_tc_ui.hb_runtime import HBRuntime
 
ALPHA = 0.5
 
def collect_args() -> argparse.Namespace:
    """Set command line arguments"""
    parser = argparse.ArgumentParser()
    parser.add_argument('--config', help="Config file", type=str, default="params.yaml")
    args = parser.parse_args()
    return args
 
class Tupperware(dict):
    MARKER = object()
 
    def __init__(self, value=None):
        if value is None:
            pass
        elif isinstance(value, dict):
            for key in value:
                self.__setitem__(key, value[key])
        else:
            raise TypeError('expected dict')
 
    def __setitem__(self, key, value):
        if isinstance(value, dict) and not isinstance(value, Tupperware):
            value = Tupperware(value)
        super(Tupperware, self).__setitem__(key, value)
 
    def __getitem__(self, key):
        found = self.get(key, Tupperware.MARKER)
        if found is Tupperware.MARKER:
            found = Tupperware()
            super(Tupperware, self).__setitem__(key, found)
        return found
 
    __setattr__, __getattr__ = __setitem__, __getitem__
 
def collect_tupperware() -> Tupperware:
    config = collect_args()
    params = yaml.safe_load(open(config.config))
    args = Tupperware(params)
    print(args)
    return args
 
def pre_image_op(args, index, frame_index, cam_side):
    total_car1_images = 6054
    cropped_coords = dict(Car1=dict(FV=(114, 110, 1176, 610),
                                    MVL=(343, 5, 1088, 411),
                                    MVR=(185, 5, 915, 425),
                                    RV=(186, 203, 1105, 630)),
                          Car2=dict(FV=(160, 272, 1030, 677),
                                    MVL=(327, 7, 1096, 410),
                                    MVR=(175, 4, 935, 404),
                                    RV=(285, 187, 1000, 572)))
    if args.crop:
        if int(frame_index[1:]) < total_car1_images:
            cropped_coords = cropped_coords["Car1"][cam_side]
        else:
            cropped_coords = cropped_coords["Car2"][cam_side]
    else:
        cropped_coords = None
 
    cropped_image = get_image(args, index, cropped_coords, frame_index, cam_side)
    resized_image = cv2.resize(np.array(cropped_image), (args.input_width, args.input_height),
                               cv2.INTER_LANCZOS4).transpose((2, 0, 1))
    resized_image = np.expand_dims(resized_image, axis=0).astype(np.float32)
    return resized_image / 255

def get_image(args, index, cropped_coords, frame_index, cam_side):
    recording_folder = "rgb_images" if index == 0 else "previous_images"
    file = f"{frame_index}_{cam_side}.png" if index == 0 else f"{frame_index}_{cam_side}_prev.png"
  
    path = os.path.join(args.dataset_dir, recording_folder, file)
   
    image = Image.open(path).convert('RGB')
   
    if args.crop:
        return image.crop(cropped_coords)
    return image
 
i = 0
def verify_onnx_model(args):
 
    image_paths = [line.rstrip('\n') for line in open("./ori_dataset/val.txt")]
    print(image_paths)
    i = 0
    for path in image_paths:
        
        frame_index, cam_side = path.split('.')[0].split('_')
        previous_frame = pre_image_op(args, -1, frame_index, cam_side)
        
        current_frame = pre_image_op(args, 0, frame_index, cam_side)
        np.save(f"calibration_data_rgb/previous_data_npy/previous_index{i}.npy",previous_frame)
        np.save(f"calibration_data_rgb/rgb_data_npy/current_index{i}.npy",current_frame)
        i += 1
 
if __name__ == "__main__":
    # load your predefined ONNX model
    args = collect_tupperware()
    verify_onnx_model(args)eszx

数据预处理流程要和训练集合数据处理流程一样。

三、配置文件设置

calibration_parameters:
  cal_data_dir: ./calibration_data_rgb/previous_data_npy;./calibration_data_rgb/rgb_data_npy
  cal_data_type: ''
  calibration_type: default
  optimization: ''
  per_channel: false
  quant_config: ''
  run_on_bpu: ''
  run_on_cpu: ''
compiler_parameters:
  advice: 0
  balance_factor: 0
  compile_mode: latency
  core_num: 1
  debug: true
  jobs: 16
  max_time_per_fc: 0
  optimize_level: O2
input_parameters:
  input_layout_rt: ''
  input_layout_train: NCHW;NCHW
  input_name: input.1;input.55
  input_shape: 1x3x288x544;1x3x288x544
  input_type_rt: featuremap;featuremap
  input_type_train: featuremap;featuremap
  # mean_value: 0;0
  #norm_type: data_mean_and_scale;data_mean_and_scale
  #scale_value: 0.003921568627451;0.003921568627451
  separate_batch: false
model_parameters:
  debug_mode: ''
  layer_out_dump: false
  march: nash-e
  node_info: ''
  onnx_model: omnidet_float32_opset12.onnx
  output_model_file_prefix: omnidet_float32_opset12
  output_nodes: ''
  remove_node_name: ''
  remove_node_type: ''
  working_dir: ./model_output

因为是双输入，所以 cal_data_dir: ./calibration_data_rgb/previous_data_npy;./calibration_data_rgb/rgb_data_npy 这里配置两个校准集路径，后面的一些配置都配置成双份的。

四、模型量化编译

hb_compile --config config.file

编译时间较长，耐心等待即可。

在 model_output 中我们会得到 hbm 模型文件，该模型文件是经过量化之后的。

五、板端编译部署

可以选择在服务器端进行编译，也可以在板端进行编译。NVCC 交叉编译。前提是需要安装好 gcc，g++，可以用 which gcc，which g++查看是否安装了 C++编译器。我们以 Cmake 形式进行工程编译。

CMakeList.txt 如下：

# CMakeLists.txt
cmake_minimum_required(VERSION 3.0)
 
project(test_sample)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 -Wl,-unresolved-symbols=ignore-in-shared-libs")
 
message(STATUS "Build type: ${CMAKE_BUILD_TYPE}")
 
set(CMAKE_CXX_FLAGS_DEBUG "-g -O0")
set(CMAKE_C_FLAGS_DEBUG "-g -O0")
set(CMAKE_CXX_FLAGS_RELEASE " -O3 ")
set(CMAKE_C_FLAGS_RELEASE " -O3 ")
 
set(CMAKE_BUILD_TYPE ${build_type})
 
set(DEPS_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/deps_aarch64)
 
include_directories(${DEPS_ROOT}/ucp/include)
include_directories(/map/haiquan.liu/test2/test_sample/deps_aarch64/opencv/include)
 
link_directories(${DEPS_ROOT}/ucp/lib)
# 设置 OpenCV 头文件和库路径
add_executable(run_sample src/main.cc)
target_link_libraries(run_sample dnn hbucp)
target_link_libraries(run_sample "/map/haiquan.liu/test2/test_sample/deps_aarch64/opencv/lib/libopencv_world.so")
main函数如下：
#include 
#include 
#include 
#include 

#include 

#include "hobot/dnn/hb_dnn.h"
#include "hobot/hb_ucp.h"
#include "hobot/hb_ucp_sys.h"
using namespace cv;

#define ALIGN(value, alignment) (((value) + ((alignment)-1)) & ~((alignment)-1))
#define ALIGN_32(value) ALIGN(value, 32)

const char* hbm_path = "omnidet_float32_opset12.hbm";
std::string data_path1 = "input_file1.bin";
std::string data_path2 = "input_file2.bin";

// Read binary input file
int read_binary_file(std::string file_path, char **bin, int *length) {
    std::ifstream ifs(file_path, std::ios::in | std::ios::binary);
    ifs.seekg(0, std::ios::end);
    *length = ifs.tellg();
    ifs.seekg(0, std::ios::beg);
    *bin = new char[sizeof(char) * (*length)];
    ifs.read(*bin, *length);
    ifs.close();
    return 0;
}

// Prepare input tensor and output tensor
int prepare_tensor(hbDNNTensor *input_tensor, hbDNNTensor *output_tensor,
                   hbDNNHandle_t dnn_handle) {
  // Get input and output tensor counts
  int input_count = 0;
  int output_count = 0;
  hbDNNGetInputCount(&input_count, dnn_handle);
  hbDNNGetOutputCount(&output_count, dnn_handle);

  hbDNNTensor *input = input_tensor;
  // Get the properties of the input tensor
  for (int i = 0; i < input_count; i++) {
    hbDNNGetInputTensorProperties(&input[i].properties, dnn_handle, i);
    // Calculate the stride of the input tensor
    auto dim_len = input[i].properties.validShape.numDimensions;
    for (int32_t dim_i = dim_len - 1; dim_i >= 0; --dim_i) {
      if (input[i].properties.stride[dim_i] == -1) {
        auto cur_stride =
            input[i].properties.stride[dim_i + 1] *
            input[i].properties.validShape.dimensionSize[dim_i + 1];
        input[i].properties.stride[dim_i] = ALIGN_32(cur_stride);
      }
    }
    // Calculate the memory size of the input tensor and allocate cache memory
    int input_memSize = input[i].properties.stride[0] *
                        input[i].properties.validShape.dimensionSize[0];
    hbUCPMallocCached(&input[i].sysMem, input_memSize, 0);
  }

  hbDNNTensor *output = output_tensor;
  // Get the properties of the input tensor
  for (int i = 0; i < output_count; i++) {
    hbDNNGetOutputTensorProperties(&output[i].properties, dnn_handle, i);
    // Calculate the memory size of the output tensor and allocate cache memory
    int output_memSize = output[i].properties.alignedByteSize;
    hbUCPMallocCached(&output[i].sysMem, output_memSize, 0);

    // Show how to get output name
    const char *output_name;
    hbDNNGetOutputName(&output_name, dnn_handle, i);
  }
  return 0;
}

int main() {
    // 获取模型句柄
    hbDNNPackedHandle_t packed_dnn_handle;
    hbDNNHandle_t dnn_handle;
    hbDNNInitializeFromFiles(&packed_dnn_handle, &hbm_path, 1);
    const char **model_name_list;
    int model_count = 0;
    hbDNNGetModelNameList(&model_name_list, &model_count, packed_dnn_handle);
    hbDNNGetModelHandle(&dnn_handle, packed_dnn_handle, model_name_list[0]);

        // Prepare input and output tensor
    std::vector<hbDNNTensor> input_tensors;
    std::vector<hbDNNTensor> output_tensors;
    int input_count = 0;
    int output_count = 0;
    hbDNNGetInputCount(&input_count, dnn_handle);
    hbDNNGetOutputCount(&output_count, dnn_handle);
    input_tensors.resize(input_count);
    output_tensors.resize(output_count);

    // Initialize and malloc the tensor
    prepare_tensor(input_tensors.data(), output_tensors.data(), dnn_handle);

    // 复制输入数据到输入张量
    int32_t data_length1 = 0;
    int32_t data_length2 = 0;
    char *data1 = nullptr, *data2 = nullptr;

    // 读取两个输入数据
    auto ret1 = read_binary_file(data_path1, &data1, &data_length1);
    auto ret2 = read_binary_file(data_path2, &data2, &data_length2);

    // 将数据复制到输入张量
    memcpy(reinterpret_cast<char *>(input_tensors[0].sysMem.virAddr), data1, input_tensors[0].sysMem.memSize);
    memcpy(reinterpret_cast<char *>(input_tensors[1].sysMem.virAddr), data2, input_tensors[1].sysMem.memSize);

    // 刷新内存，确保数据写入
    hbUCPMemFlush(&(input_tensors[0].sysMem), HB_SYS_MEM_CACHE_CLEAN);
    hbUCPMemFlush(&(input_tensors[1].sysMem), HB_SYS_MEM_CACHE_CLEAN);

    // 提交推理任务并等待完成
    hbUCPTaskHandle_t task_handle{nullptr};
    hbDNNTensor *output = output_tensors.data();

    hbDNNInferV2(&task_handle, output, input_tensors.data(), dnn_handle);

    // 等待任务完成
    hbUCPSchedParam ctrl_param;
    HB_UCP_INITIALIZE_SCHED_PARAM(&ctrl_param);
    ctrl_param.backend = HB_UCP_BPU_CORE_ANY;
    hbUCPSubmitTask(task_handle, &ctrl_param);
    hbUCPWaitTaskDone(task_handle, 0);

    // 解析推理结果并处理每个输出张量
    for (int i = 0; i < 4; ++i) {
        hbUCPMemFlush(&output_tensors[i].sysMem, HB_SYS_MEM_CACHE_INVALIDATE);
        auto result = reinterpret_cast<float *>(output_tensors[i].sysMem.virAddr);

        // 处理每个任务的结果
        if (i == 0) {
            // 任务1: 处理 output_tensors[0]
            // 例如分类任务，分割任务等
        } else if (i == 1) {
            // 任务2: 处理 output_tensors[1]
            // int height = output_tensors[i].properties.shape[1]; // 输出图像的高度
            // int width = output_tensors[i].properties.shape[2];  // 输出图像的宽度
            
            int height = 288; // 输出图像的高度
            int width = 544;  // 输出图像的宽度

            // 假设result是一个(float类型的)指向概率的指针，维度为(10, 288, 544)
            // 10个类别，对应每个像素的类别概率
            float* result = reinterpret_cast<float*>(output_tensors[i].sysMem.virAddr);

            // 创建一个伪彩色图像用于渲染（将类别索引映射到颜色）
            cv::Mat rendered_image(height, width, CV_8UC3);

            // 创建一个颜色映射（例如，10个类别对应不同颜色）
            std::vector<cv::Vec3b> color_map = {
                cv::Vec3b(0, 0, 255),   // 类别 0 (红色)
                cv::Vec3b(0, 255, 0),   // 类别 1 (绿色)
                cv::Vec3b(255, 0, 0),   // 类别 2 (蓝色)
                cv::Vec3b(0, 255, 255), // 类别 3 (青色)
                cv::Vec3b(255, 255, 0), // 类别 4 (黄色)
                cv::Vec3b(255, 0, 255), // 类别 5 (品红)
                cv::Vec3b(128, 128, 128), // 类别 6 (灰色)
                cv::Vec3b(255, 165, 0), // 类别 7 (橙色)
                cv::Vec3b(255, 20, 147), // 类别 8 (深粉色)
                cv::Vec3b(0, 191, 255)  // 类别 9 (深天蓝)
            };

            // 对每个像素进行argmax操作，选择概率最大类别
            for (int h = 0; h < height; ++h) {
                for (int w = 0; w < width; ++w) {
                    // 每个像素有10个类别的概率，找到最大概率对应的类别索引
                    int class_id = 0;
                    float max_prob = result[class_id * height * width + h * width + w];
                    for (int c = 1; c < 10; ++c) {
                        float prob = result[c * height * width + h * width + w];
                        if (prob > max_prob) {
                            max_prob = prob;
                            class_id = c;
                        }
                    }

                    // 将类别索引映射到颜色
                    rendered_image.at<cv::Vec3b>(h, w) = color_map[class_id];
                }
            }
            // 保存渲染图像
            cv::imwrite("segmentation_output.png", rendered_image);

        } else if (i == 2) {
            // 任务3: 处理 output_tensors[2]
        } else if (i == 3) {
            // 任务4: 处理 output_tensors[3]
        }
    }

    return 0;
}

注意将环境以来迁移过来 ，如在服务器的 docker 环境下，则不必迁移过来。直接在服务器端编译即可。

布局如下

然后 mkdir build，cd build，cmake ..，make 依次操作。

会在 build 文件夹下得到编译过的二进制文件，将输入数据以及模型和二进制文件迁移到同个文件夹下面。运行。/二进制文件即可。分割效果图如下