yolov8在旭日x3派板子上的部署
主要内容如下：
- 用修改后的yolov8训练自己的数据集
- Pytorch到ONNX模型转换
- ONNX到BIN转换
- 旭日x3派上的实时检测
以下是官方对于yolov8的介绍和链接
[Ultralytics YOLOv8](https://github.com/ultralytics/ultralytics), developed by [Ultralytics](https://ultralytics.com), is a cutting-edge, state-of-the-art (SOTA) model that builds upon the success of previous YOLO versions and introduces new features and improvements to further boost performance and flexibility. YOLOv8 is designed to be fast, accurate, and easy to use, making it an excellent choice for a wide range of object detection, image segmentation and image classification tasks. To request an Enterprise License please complete the form at [Ultralytics Licensing](https://ultralytics.com/license).
其中yolov8n比yolov5s参数量少一倍的情况下，精度可以持平yolov5s，同时与yolov6和yolov7相比也是更胜一筹。
get started
git clone https://github.com/Hzbupahaozi/yolov8_xj3_deploy.git
cd yolov8_xj3_deploy
pip install -r requirements.txt
python setup.py installtrain
旭日x3派开发板的亮点在于其BPU 5T的int8算力\ BPU的高效源于其特别的网络结构————可变组卷积（VarGNet），具体可以参考https://arxiv.org/abs/1907.05653\ 因此对原本的yolov8网络结构进行修改，即将yolov8.yaml修改为x3pi_model_config/yolov8-vargnetct.yaml\ 准备好yolo格式的数据集后就可以开始训练了
yolo task=detect mode=train model=./x3pi_config/yolov8n-vargnetct.yaml data=./mydata.yaml batch=32 epochs=80 imgsz=640 workers=8 device=0训练结果如下图
pytorch模型到onnx模型的转换
以下是三木大佬的原话博客：\ 由于 YOLOv8 原始仓库转换的模型把预测特征解码放到了模型里面，这其实会导致两个问题：
- 目标检测模型并不是要计算所有框并且回归的，通过置信度可以筛选超过 90% 的框，这些都不需要额外计算的，因此我提供了一个传参 x3pi=True 用来删除解码等操作。
- 原始模型的解码操作 BPU 不能很好的加速，那索性就全用 CPU 处理好了。
转换命令：
yolo export detect model=best.pt format=onnx x3pi=Trueonnx到bin的转换
这里使用docker挂载天工开物的开发工具包实现\ 教程参考：BPU部署走出新手村\ 需要注意的是要下载最新的2.5.2的容器和2.5.2的oe文档，上述教程使用版本过旧（导致我搞了好一段时间没啥进展），
docker run -it --rm -v "D:\docker\horizon_xj3_open_explorer_v2.5.2-py38_20230331":/open_explorer -v "D:\docker\BPUCodes":/data/horizon_x3/codes openexplorer/ai_toolchain_ubuntu_20_xj3_cpu:v2.5.2
cd /open_explorer/ddk/samples/ai_toolchain/horizon_model_convert_sample/04_detection/ mkdir ./yolov8/ cd ./yolov8/
为了模型的轻量化在转换为bin模型时会将模型参数通过PTQ转换为int8，需要准备校准数据集，具体脚本为make_calib.py\
然后把一些相关文件（onnx模型，转换文件my_config.yaml，校准数据集文件夹）拷贝进去
```sh
docker cp 本地文件路径 容器:/open_explorer/ddk/samples/ai_toolchain/horizon_model_convert_sample/04_detection/yolov8/其中容器可通过以下命令查看
docker ps然后就可以开始转换了
# check
hb_mapper checker --model-type onnx \
--model best.onnx \
--march bernoulli2
# export
hb_mapper makertbin --config my_config.yaml --model-type onnx然后就可以看到很多日志
==============================================================================================================================================
Node ON Subgraph Type Cosine Similarity Threshold In/Out DataType
----------------------------------------------------------------------------------------------------------------------------------------------
HZ_PREPROCESS_FOR_images BPU id(0) HzSQuantizedPreprocess 0.999949 127.000000 int8/int8
Conv_0 BPU id(0) HzSQuantizedConv 0.999869 1.502627 int8/int8
Conv_2 BPU id(0) HzSQuantizedConv 0.999687 17.772688 int8/int8
Conv_4 BPU id(0) HzSQuantizedConv 0.999460 16.475630 int8/int8
Split_6 BPU id(0) Split int8/int8
Conv_7 BPU id(0) HzSQuantizedConv 0.999311 14.938033 int8/int8
Conv_9 BPU id(0) HzSQuantizedConv 0.999322 16.493052 int8/int8
Conv_11 BPU id(0) HzSQuantizedConv 0.999112 19.128515 int8/int8
Conv_13 BPU id(0) HzSQuantizedConv 0.999251 18.809891 int8/int8
UNIT_CONV_FOR_Add_15 BPU id(0) HzSQuantizedConv 0.999442 14.938033 int8/int8
...CONV_FOR_onnx::Concat_444_0.14999_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
UNIT_CONV_FOR_input.24_0.14999_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
Concat_16 BPU id(0) Concat 0.999392 14.938033 int8/int8
Conv_17 BPU id(0) HzSQuantizedConv 0.999088 19.048702 int8/int8
Conv_19 BPU id(0) HzSQuantizedConv 0.999166 11.896363 int8/int8
Conv_21 BPU id(0) HzSQuantizedConv 0.999127 13.429939 int8/int8
Split_23 BPU id(0) Split int8/int8
Conv_24 BPU id(0) HzSQuantizedConv 0.996399 13.270452 int8/int8
Conv_26 BPU id(0) HzSQuantizedConv 0.996832 9.714583 int8/int8
Conv_28 BPU id(0) HzSQuantizedConv 0.995260 6.936894 int8/int8
Conv_30 BPU id(0) HzSQuantizedConv 0.996167 5.460835 int8/int8
UNIT_CONV_FOR_Add_32 BPU id(0) HzSQuantizedConv 0.998495 13.270452 int8/int8
Conv_33 BPU id(0) HzSQuantizedConv 0.996459 8.126689 int8/int8
Conv_35 BPU id(0) HzSQuantizedConv 0.996397 5.057761 int8/int8
Conv_37 BPU id(0) HzSQuantizedConv 0.995583 4.751214 int8/int8
Conv_39 BPU id(0) HzSQuantizedConv 0.995153 6.242482 int8/int8
UNIT_CONV_FOR_Add_41 BPU id(0) HzSQuantizedConv 0.998127 8.126689 int8/int8
...CONV_FOR_onnx::Concat_469_0.09842_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
UNIT_CONV_FOR_input.88_0.09842_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
UNIT_CONV_FOR_input.124_0.09842_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
Concat_42 BPU id(0) Concat 0.998541 13.270452 int8/int8
Conv_43 BPU id(0) HzSQuantizedConv 0.997992 12.499084 int8/int8
Conv_45 BPU id(0) HzSQuantizedConv 0.999056 6.938571 int8/int8
Conv_47 BPU id(0) HzSQuantizedConv 0.999112 6.857578 int8/int8
Split_49 BPU id(0) Split int8/int8
Conv_50 BPU id(0) HzSQuantizedConv 0.998487 6.594507 int8/int8
Conv_52 BPU id(0) HzSQuantizedConv 0.998260 6.571420 int8/int8
Conv_54 BPU id(0) HzSQuantizedConv 0.997727 4.126614 int8/int8
Conv_56 BPU id(0) HzSQuantizedConv 0.996096 4.693301 int8/int8
UNIT_CONV_FOR_Add_58 BPU id(0) HzSQuantizedConv 0.998729 6.594507 int8/int8
Conv_59 BPU id(0) HzSQuantizedConv 0.997990 6.840492 int8/int8
Conv_61 BPU id(0) HzSQuantizedConv 0.997279 4.396142 int8/int8
Conv_63 BPU id(0) HzSQuantizedConv 0.996712 3.522115 int8/int8
Conv_65 BPU id(0) HzSQuantizedConv 0.995933 4.622054 int8/int8
UNIT_CONV_FOR_Add_67 BPU id(0) HzSQuantizedConv 0.998451 6.840492 int8/int8
...CONV_FOR_onnx::Concat_507_0.05330_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
UNIT_CONV_FOR_input.188_0.05330_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
UNIT_CONV_FOR_input.224_0.05330_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
Concat_68 BPU id(0) Concat 0.998793 6.594507 int8/int8
Conv_69 BPU id(0) HzSQuantizedConv 0.998581 6.768975 int8/int8
Conv_71 BPU id(0) HzSQuantizedConv 0.999188 4.436905 int8/int8
Conv_73 BPU id(0) HzSQuantizedConv 0.999031 4.389203 int8/int8
Split_75 BPU id(0) Split int8/int8
Conv_76 BPU id(0) HzSQuantizedConv 0.998980 4.278706 int8/int8
Conv_78 BPU id(0) HzSQuantizedConv 0.998558 5.475333 int8/int8
Conv_80 BPU id(0) HzSQuantizedConv 0.998531 3.166869 int8/int8
Conv_82 BPU id(0) HzSQuantizedConv 0.998872 4.244014 int8/int8
UNIT_CONV_FOR_Add_84 BPU id(0) HzSQuantizedConv 0.999180 4.278706 int8/int8
...CONV_FOR_onnx::Concat_545_0.03857_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
UNIT_CONV_FOR_input.288_0.03857_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
Concat_85 BPU id(0) Concat 0.999096 4.278706 int8/int8
Conv_86 BPU id(0) HzSQuantizedConv 0.998955 4.898155 int8/int8
Conv_88 BPU id(0) HzSQuantizedConv 0.998856 4.396799 int8/int8
MaxPool_90 BPU id(0) HzQuantizedMaxPool 0.999222 6.915278 int8/int8
MaxPool_91 BPU id(0) HzQuantizedMaxPool 0.999414 6.915278 int8/int8
MaxPool_92 BPU id(0) HzQuantizedMaxPool 0.999525 6.915278 int8/int8
Concat_93 BPU id(0) Concat 0.999353 6.915278 int8/int8
Conv_94 BPU id(0) HzSQuantizedConv 0.999698 6.915278 int8/int8
ConvTranspose_96 BPU id(0) HzSQuantizedConvTranspose 0.999779 3.459090 int8/int8
UNIT_CONV_FOR_onnx::Conv_538_0.04263_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
Concat_98 BPU id(0) Concat 0.999320 5.414181 int8/int8
Conv_99 BPU id(0) HzSQuantizedConv 0.999117 5.414181 int8/int8
Split_101 BPU id(0) Split int8/int8
Conv_102 BPU id(0) HzSQuantizedConv 0.998763 5.000547 int8/int8
Conv_104 BPU id(0) HzSQuantizedConv 0.998287 5.897664 int8/int8
Conv_106 BPU id(0) HzSQuantizedConv 0.997791 4.177810 int8/int8
Conv_108 BPU id(0) HzSQuantizedConv 0.997375 4.131814 int8/int8
...CONV_FOR_onnx::Concat_580_0.03904_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
UNIT_CONV_FOR_input.372_0.03904_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
Concat_110 BPU id(0) Concat 0.998700 5.000547 int8/int8
Conv_111 BPU id(0) HzSQuantizedConv 0.998120 4.958530 int8/int8
ConvTranspose_113 BPU id(0) HzSQuantizedConvTranspose 0.998221 5.232376 int8/int8
UNIT_CONV_FOR_onnx::Conv_500_0.05410_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
Concat_115 BPU id(0) Concat 0.998037 6.871005 int8/int8
Conv_116 BPU id(0) HzSQuantizedConv 0.997310 6.871005 int8/int8
Split_118 BPU id(0) Split int8/int8
Conv_119 BPU id(0) HzSQuantizedConv 0.996089 5.632505 int8/int8
Conv_121 BPU id(0) HzSQuantizedConv 0.995166 5.500845 int8/int8
Conv_123 BPU id(0) HzSQuantizedConv 0.993983 6.010270 int8/int8
Conv_125 BPU id(0) HzSQuantizedConv 0.993902 7.229157 int8/int8
...CONV_FOR_onnx::Concat_604_0.04773_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
UNIT_CONV_FOR_input.436_0.04773_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
Concat_127 BPU id(0) Concat 0.996177 5.632505 int8/int8
Conv_128 BPU id(0) HzSQuantizedConv 0.994840 6.062278 int8/int8
Conv_130 BPU id(0) HzSQuantizedConv 0.996329 7.191900 int8/int8
...R_onnx::ConvTranspose_597_0.05140_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
Concat_132 BPU id(0) Concat 0.997258 6.528210 int8/int8
Conv_133 BPU id(0) HzSQuantizedConv 0.996429 6.528210 int8/int8
Split_135 BPU id(0) Split int8/int8
Conv_136 BPU id(0) HzSQuantizedConv 0.994795 6.808836 int8/int8
Conv_138 BPU id(0) HzSQuantizedConv 0.994417 6.371107 int8/int8
Conv_140 BPU id(0) HzSQuantizedConv 0.993589 6.077781 int8/int8
Conv_142 BPU id(0) HzSQuantizedConv 0.993910 6.707007 int8/int8
...CONV_FOR_onnx::Concat_629_0.05270_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
UNIT_CONV_FOR_input.504_0.05270_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
Concat_144 BPU id(0) Concat 0.995724 6.808836 int8/int8
Conv_145 BPU id(0) HzSQuantizedConv 0.995444 6.693432 int8/int8
Conv_147 BPU id(0) HzSQuantizedConv 0.993256 6.435788 int8/int8
...R_onnx::ConvTranspose_573_0.03540_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
Concat_149 BPU id(0) Concat 0.998485 4.496263 int8/int8
Conv_150 BPU id(0) HzSQuantizedConv 0.997724 4.496263 int8/int8
Split_152 BPU id(0) Split int8/int8
Conv_153 BPU id(0) HzSQuantizedConv 0.997050 4.260736 int8/int8
Conv_155 BPU id(0) HzSQuantizedConv 0.996641 5.090493 int8/int8
Conv_157 BPU id(0) HzSQuantizedConv 0.994862 3.914743 int8/int8
Conv_159 BPU id(0) HzSQuantizedConv 0.992631 4.832358 int8/int8
...CONV_FOR_onnx::Concat_654_0.03318_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
UNIT_CONV_FOR_input.572_0.03318_TO_FUSE_SCALE BPU id(0) HzSQuantizedConv int8/int8
Concat_161 BPU id(0) Concat 0.995999 4.260736 int8/int8
Conv_162 BPU id(0) HzSQuantizedConv 0.995250 4.214143 int8/int8
Conv_164 BPU id(0) HzSQuantizedConv 0.995858 7.191900 int8/int8
Mul_166 BPU id(0) HzLut 0.996182 7.299980 int8/int8
Conv_167 BPU id(0) HzSQuantizedConv 0.997211 7.295052 int8/int8
Mul_169 BPU id(0) HzLut 0.997212 7.081635 int8/int8
Conv_170 BPU id(0) HzSQuantizedConv 0.998338 7.075689 int8/int32
Conv_172 BPU id(0) HzSQuantizedConv 0.995082 7.191900 int8/int8
Mul_174 BPU id(0) HzLut 0.993972 6.291241 int8/int8
Conv_175 BPU id(0) HzSQuantizedConv 0.994323 6.279607 int8/int8
Mul_177 BPU id(0) HzLut 0.994310 10.864396 int8/int8
Conv_178 BPU id(0) HzSQuantizedConv 0.999579 10.864189 int8/int32
Conv_180 BPU id(0) HzSQuantizedConv 0.996984 6.435788 int8/int8
Mul_182 BPU id(0) HzLut 0.996698 6.473208 int8/int8
Conv_183 BPU id(0) HzSQuantizedConv 0.997236 6.463228 int8/int8
Mul_185 BPU id(0) HzLut 0.996391 7.075638 int8/int8
Conv_186 BPU id(0) HzSQuantizedConv 0.998851 7.069662 int8/int32
Conv_188 BPU id(0) HzSQuantizedConv 0.995746 6.435788 int8/int8
Mul_190 BPU id(0) HzLut 0.995045 5.152664 int8/int8
Conv_191 BPU id(0) HzSQuantizedConv 0.995736 5.123033 int8/int8
Mul_193 BPU id(0) HzLut 0.995775 6.424686 int8/int8
Conv_194 BPU id(0) HzSQuantizedConv 0.999827 6.414289 int8/int32
Conv_196 BPU id(0) HzSQuantizedConv 0.994093 3.806277 int8/int8
Mul_198 BPU id(0) HzLut 0.993954 4.152104 int8/int8
Conv_199 BPU id(0) HzSQuantizedConv 0.994331 4.087798 int8/int8
Mul_201 BPU id(0) HzLut 0.994222 4.594001 int8/int8
Conv_202 BPU id(0) HzSQuantizedConv 0.999736 4.548010 int8/int32
Conv_204 BPU id(0) HzSQuantizedConv 0.998809 3.806277 int8/int8
Mul_206 BPU id(0) HzLut 0.998360 3.759366 int8/int8
Conv_207 BPU id(0) HzSQuantizedConv 0.999285 3.673772 int8/int8
Mul_209 BPU id(0) HzLut 0.999141 4.446995 int8/int8
Conv_210 BPU id(0) HzSQuantizedConv 0.999996 4.395507 int8/int32
2023-04-30 12:46:11,222 INFO [Sun Apr 30 12:46:11 2023] End to Horizon NN Model Convert.
2023-04-30 12:46:11,268 INFO start convert to *.bin file....
2023-04-30 12:46:11,299 INFO ONNX model output num : 6
2023-04-30 12:46:11,302 INFO ############# model deps info #############
2023-04-30 12:46:11,303 INFO hb_mapper version : 1.15.5
2023-04-30 12:46:11,303 INFO hbdk version : 3.44.1
2023-04-30 12:46:11,303 INFO hbdk runtime version: 3.15.17.0
2023-04-30 12:46:11,303 INFO horizon_nn version : 0.16.3
2023-04-30 12:46:11,304 INFO ############# model_parameters info #############
2023-04-30 12:46:11,304 INFO onnx_model : /open_explorer/ddk/samples/ai_toolchain/horizon_model_convert_sample/04_detection/myyolov8/best.onnx
2023-04-30 12:46:11,304 INFO BPU march : bernoulli2
2023-04-30 12:46:11,304 INFO layer_out_dump : False
2023-04-30 12:46:11,305 INFO log_level : DEBUG
2023-04-30 12:46:11,305 INFO working dir : /open_explorer/ddk/samples/ai_toolchain/horizon_model_convert_sample/04_detection/myyolov8/model_outputs
2023-04-30 12:46:11,305 INFO output_model_file_prefix: yolov8_horizon
2023-04-30 12:46:11,305 INFO ############# input_parameters info #############
2023-04-30 12:46:11,305 INFO ------------------------------------------
2023-04-30 12:46:11,306 INFO ---------input info : images ---------
2023-04-30 12:46:11,306 INFO input_name : images
2023-04-30 12:46:11,306 INFO input_type_rt : nv12
2023-04-30 12:46:11,306 INFO input_space&range : regular
2023-04-30 12:46:11,306 INFO input_layout_rt : NHWC
2023-04-30 12:46:11,307 INFO input_type_train : rgb
2023-04-30 12:46:11,307 INFO input_layout_train : NCHW
2023-04-30 12:46:11,307 INFO norm_type : data_scale
2023-04-30 12:46:11,307 INFO input_shape : 1x3x640x640
2023-04-30 12:46:11,307 INFO input_batch : 1
2023-04-30 12:46:11,308 INFO scale_value : 0.003921568627451,
2023-04-30 12:46:11,308 INFO cal_data_dir : /open_explorer/ddk/samples/ai_toolchain/horizon_model_convert_sample/04_detection/myyolov8/calib_640_f32
2023-04-30 12:46:11,308 INFO cal_data_type : float32
2023-04-30 12:46:11,308 INFO ---------input info : images end -------
2023-04-30 12:46:11,309 INFO ------------------------------------------
2023-04-30 12:46:11,309 INFO ############# calibration_parameters info #############
2023-04-30 12:46:11,309 INFO preprocess_on : False
2023-04-30 12:46:11,309 INFO calibration_type: : default
2023-04-30 12:46:11,309 INFO ############# compiler_parameters info #############
2023-04-30 12:46:11,310 INFO hbdk_pass_through_params: --O3 --core-num 2 --fast
2023-04-30 12:46:11,310 INFO input-source : {'images': 'pyramid', '_default_value': 'ddr'}
2023-04-30 12:46:11,316 INFO Convert to runtime bin file successfully!
2023-04-30 12:46:11,316 INFO End Model Convert我们可以看到所有算子都是在BPU上跑了，所以可以起到加速推理的作用，同时也说明对yolov8模型的修改是有效的！因为我尝试过使用最原本的yolov8n结构，结果还是有一部分不能在bpu上跑\ 模型的输出结果在model_output中，其中的yolov8_horizon.bin就是我们需要的模型
最后一步 上板！
首先我们可以测试以下地平线预装模型yolov5_672x672_nv12.bin，yolov5s_672x672_nv12.bin的效果
hrt_model_exec perf \
--model_file yolov5s_672x672_nv12.bin \
--model_name="" \
--core_id=0 \
--frame_count=500 \
--perf_time=0 \
--thread_num=1 \
--profile_path="."
hrt_model_exec perf \
--model_file fcos_512x512_nv12.bin \
--model_name="" \
--core_id=0 \
--frame_count=500 \
--perf_time=0 \
--thread_num=1 \
--profile_path="."结果如下：
Running condition:
Thread number is: 1
Frame count is: 500
Program run time: 36346.591000 ms
Perf result:
Frame totally latency is: 36325.519531 ms
Average latency is: 72.651039 ms
Frame rate is: 13.756448 FPS
Running condition:
Thread number is: 1
Frame count is: 500
Program run time: 36308.297000 ms
Perf result:
Frame totally latency is: 36286.542969 ms
Average latency is: 72.573090 ms
Frame rate is: 13.770957 FPS然后就是我们模型的效果
hrt_model_exec perf \
--model_file yolov8s-vargnetct-det.bin \
--model_name="" \
--core_id=0 \
--frame_count=500 \
--perf_time=0 \
--thread_num=1 \
--profile_path="."日志如下：
Running condition:
Thread number is: 1
Frame count is: 500
Program run time: 20303.004000 ms
Perf result:
Frame totally latency is: 20181.697266 ms
Average latency is: 40.36.3396 ms
Frame rate is: 24.626898 FPS强！ 最后就是调用usb摄像头进行实时检测，脚本为camera.py\ 代码中 72-144 行是解析模型输出的关键内容，由于这部分在 CPU 跑，因此我主要使用 numpy 高效并行优化了一下后处理中的框回归计算，然后使用 opencv-python 提供的 NMS 接口对多余重复框去重复\ 代码中 170-176 首先推理了 10 次空图做了一下预热，让模型运行更稳定\ 代码中 199-202 打印了推理过程中，除了画框部分的耗时情况\ 执行代码后，会打印如下日志：
\
前处理耗时：7 ms\
模型推理： 44 ms\
后处理耗时: 8 ms\
整体耗时： 59 ms
前前后后耗时将近一个月，中间一段时间将近放弃部署直接在电脑上跑，这两天看了一些资料加班了一下最后总算是搞完了，最后感谢以下地平线论坛里的工作人员，还有最最感谢的就是三木大佬，本次参考triple-Mu/yolov8
后续再记录一下，实现了在本地的使用onnxruntime进行推理，在inference文件夹中，其中包括
- 只导出backbone+neck的分别是自制数据集和coco数据集的onnx模型的推理\
- 正常export后onnx模型的推理\
\
\ 可以看到旭日x3派经过BPU部署后的性能和笔记本的cpu推理的性能差不多
地平线官方实现的检测算法的精度以及速度记录
To Do List
- python的多线程推理
- C++封装python
- C++推理