在Yolov5 Yolov4 Yolov3 TensorRT 实现Implementation

在Yolov5 Yolov4 Yolov3 TensorRT 实现Implementation

news: yolov5 support

引论

该项目是nvidia官方yolo-tensorrt的封装实现。你必须有经过训练的yolo模型（.weights）和来自darknet（yolov3&yolov4）的.cfg文件。对于yolov5，需要Pythorch中的模型文件（yolov5s.yaml）和经过训练的权重文件（yolov5s.pt）。

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

yolov5s , yolov5m , yolov5l , yolov5x tutorial
yolov4 , yolov4-tiny
yolov3 , yolov3-tiny

Features

inequal net width and height
batch inference
support FP32,FP16,INT8
dynamic input size

PLATFORM & BENCHMARK

windows 10
ubuntu 18.04
L4T (Jetson platform)

BENCHMARK

x86 (inference time)

model	size	gpu	fp32	fp16	INT8
yolov5s	640x640	1080ti	8ms	/	7ms
yolov5m	640x640	1080ti	13ms	/	11ms
yolov5l	640x640	1080ti	20ms	/	15ms
yolov5x	640x640	1080ti	30ms	/	23ms

Jetson NX with Jetpack4.4.1 (inference / detect time)

model	size	gpu	fp32	fp16	INT8
yolov3	416x416	nx	105ms/120ms	30ms/48ms	20ms/35ms
yolov3-tiny	416x416	nx	14ms/23ms	8ms/15ms	12ms/19ms
yolov4-tiny	416x416	nx	13ms/23ms	7ms/16ms	7ms/15ms
yolov4	416x416	nx	111ms/125ms	55ms/65ms	47ms/57ms
yolov5s	416x416	nx	47ms/88ms	33ms/74ms	28ms/64ms
yolov5m	416x416	nx	110ms/145ms	63ms/101ms	49ms/91ms
yolov5l	416x416	nx	205ms/242ms	95ms/123ms	76ms/118ms
yolov5x	416x416	nx	351ms/405ms	151ms/183ms	114ms/149ms

ubuntu

model	size	gpu	fp32	fp16	INT8
yolov4	416x416	titanv	11ms/17ms	8ms/15ms	7ms/14ms
yolov5s	416x416	titanv	7ms/22ms	5ms/20ms	5ms/18ms
yolov5m	416x416	titanv	9ms/23ms	8ms/22ms	7ms/21ms
yolov5l	416x416	titanv	17ms/28ms	11ms/23ms	11ms/24ms
yolov5x	416x416	titanv	25ms/40ms	15ms/27ms	15ms/27ms

WRAPPER

Prepare the pretrained .weights and .cfg model.

Detector detector;

Config config;

std::vector<BatchResult> res;

detector.detect(vec_image, res)

Build and use yolo-trt as DLL or SO libraries

windows10

dependency : TensorRT 7.1.3.4 , cuda 11.0 , cudnn 8.0 , opencv4 , vs2015
build:

open MSVC sln/sln.sln file

dll project : the trt yolo detector dll
demo project : test of the dll

ubuntu & L4T (jetson)

The project generate the libdetector.so lib, and the sample code. If you want to use the libdetector.so lib in your own project,this cmake file perhaps could help you .

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

cd yolo-tensorrt/

mkdir build

cd build/

cmake ..

make

./yolo-trt

API

struct Config

        std::string file_model_cfg = "configs/yolov4.cfg";

        std::string file_model_weights = "configs/yolov4.weights";

        float detect_thresh = 0.9;

        ModelType net_type = YOLOV4;

        Precision inference_precison = INT8;

        int gpu_id = 0;

        std::string calibration_image_list_file_txt = "configs/calibration_images.txt";

};

class API Detector

public:

        explicit Detector();

        ~Detector();

        void init(const Config &config);

        void detect(const std::vector<cv::Mat> &mat_image,std::vector<BatchResult> &vec_batch_result);

private:

        Detector(const Detector &);

        const Detector &operator =(const Detector &);

        class Impl;

        Impl *_impl;

};

REFERENCE

人工智能芯片与自动驾驶