yolov5 tensorrt

python tensorrt:

激活函数：hard_sigmoid

https://github.com/TrojanXu/yolov5-tensorrt

https://github.com/wang-xinyu/tensorrtx

The Pytorch implementation is ultralytics/yolov5.

Currently, we support yolov5 v1.0(yolov5s only), v2.0, v3.0 and v3.1.

• For yolov5 v3.1, please visit yolov5 release v3.1, and use the latest commit of this repo.
• For yolov5 v3.0, please visit yolov5 release v3.0, and use the latest commit of this repo.
• For yolov5 v2.0, please visit yolov5 release v2.0, and checkout commit '7cd092d' of this repo.
• For yolov5 v1.0, please visit yolov5 release v1.0, and checkout commit '0504551' of this repo.

Config

• Choose the model s/m/l/x by NET macro in yolov5.cpp
• Input shape defined in yololayer.h
• Number of classes defined in yololayer.h, DO NOT FORGET TO ADAPT THIS, If using your own model
• FP16/FP32 can be selected by the macro in yolov5.cpp
• GPU id can be selected by the macro in yolov5.cpp
• NMS thresh in yolov5.cpp
• BBox confidence thresh in yolov5.cpp
• Batch size in yolov5.cpp

v2也是leakyrelu。

2.0:这个是匹配3.0的版本，用的leakyrelu,可以检测，v3.0自己训练的精度比较低

https://github.com/BaofengZan/yolov5_2.0-TensorRt

https://github.com/AIpakchoi/yolov5_tensorrt/blob/110762eea4a7a53a91bbce35f94239136db157d4/yolov5l/common.hpp

https://github.com/baituhuangyu/yolov5-tensorrt

https://github.com/Thinker-or-Dreamer/UAV-And-RobotArm/tree/master/yolov5

HardSwishLayer_TRT

https://github.com/hlld/tensorrt-yolov5

c++不知道哪个版本：

https://github.com/enazoe/yolo-tensorrt

2020.10.23 17天以前更新的，激活函数：kHARD_SIGMOID

https://github.com/wang-xinyu/tensorrtx/tree/master/yolov5

yololayer.h

修改自己类别个数：

    static constexpr int CLASS_NUM = 1;

下面是阈值参数，nms阈值参数：

  

   

    

     

    
    

     

      #define USE_FP16 // comment out this if want to use FP32
     
    
   

    

     

    
    

     

      #define DEVICE 0 // GPU id
     
    
   

    

     

    
    

     

      #define NMS_THRESH 0.4
     
    
   

    

     

    
    

     

      #define CONF_THRESH 0.3
     
    
   

    

     

    
    

     
 
     
    
   

    

     

    
    

     
 
     
    
   

    

     

    
    

     

      std::cerr << "yolov5_rt.exe -s s // serialize model to plan file" << std::endl;
     
    
   

    

     

    
    

     

      std::cerr << "yolov5_rt.exe -e s -c 0 // detect cam" << std::endl;
     
    
   

    

     

    
    

     

      std::cerr << "yolov5_rt.exe -e s -d samples // deserialize plan file and run inference" << std::endl;
     
    
  
 

后面3行是调用demo，分两步，编译和执行

问题：原版网络检测出来的框没问题

自己训练的，tensorrt检测的与pytorch检测出来的有偏差，原因还未找到。

可能是anchors的原因，但是没找到证据。

文章来源: blog.csdn.net，作者：网奇，版权归原作者所有，如需转载，请联系作者。

原文链接：blog.csdn.net/jacke121/article/details/109233756