Deepstream_Yolo

################################################################################

Copyright (c) 2021, positive666. All rights reserved.

################################################################################

0.Instruction(coming soon update )

This repository includes Yolo series project implementation based on Deepstream. build your own Deepstream project and help you familiarize yourself with the process, you can deploy any inference model!! including yolov5, yolov4, and And some model deployment projects such as OCR project based on yolo-ob detection and more will be added later.
Reference:
you can git clone and prepare your models or test convert to trained models
yolov5: https://github.com/ultralytics/yolov5
yolov4: https://github.com/Tianxiaomo/pytorch-YOLOv4.git
download its weights!

1.Make sure your environment is configured

Check Nvidia graphics driver, CUDA, CUDNN, TensorRT, Deepstream
NOTE:If you ensure that your environment is configured, you can start to create a project!

2.Convert the trained model file into an engine file

 pip install -r requirement.txt
 cd Deepstream_Project
 cd Deepstream_Yolo

Here are the two ways,you can select yolov5s,yolov5m,yolov5l,yolov5x :

1.trans_project(onnx---->trt)

2.tensorrtx project

2.0 methods_1----------trans_project(onnx---->trt):

   cd trans_project

YOLOv4 (note: torch version==1.4.0)

   cd yolov4_convert

1)darnet models--->onnx ,(your v5 and v4 need two environments,different torch version)

   python demo_darknet2onnx.py   <cfgFile>  <weightFile>  <imageFile>  <batchSize>

2)pytorch models--->onnx:

   python demo_pytorch2onnx.py  <weight_file>  <image_path>  <batch_size>  <n_classes>  <IN_IMAGE_H>  <IN_IMAGE_W>

3)onnx->trt,you will get model yolov4 trt file

   trtexec --onnx=<your onnx name >.onnx --explicitBatch --saveEngine=yolov4_1_3_320_512_fp16.engine --workspace=4096 --fp16

if you can simpilfy onnx ,make sure your onnxsim and run :

   python -m onnxsim ./weights/yolov4.onnx  ./weights/yolov4_sim.onnx

if you want to compile yolov4.cpp ,generate yolov4 trt file in your build folder:

       mkdir build && cd build
   cmake ..
   make 
       ./yolov4_convert   ../config.yaml       ../images
                        <your config file>   <your test data folder>

YOLOV5(note: torch version==1.7.0 )

   cd yolov5_convert

1)your source actvivate your python environment,input your yolov5 models into models/

2)run demo_pytorch2onnx.py to generate yolov5 onnx file:

   python demo_pytorch2onnx.py --weights ./weights/yolov5x.pt  --img 640   --batch 1       
                                           <your model path>   <imagsize>  <batchsize>

or you can try run:

   python convert.py  --weights <your modle path >  --img-size<imgsize> --bata-size<default=1>

if you can simpilfy onnx ,make sure your onnxsim and run :

   python -m onnxsim ./weights/yolov5x.onnx  ./weights/yolov5x_sim.onnx
                       <input>                     <output>

3)compile yolov5.cpp ,generate yolov5 enginel:

   mkdir build && cd build
   cmake ..
   make 
   ./yolov5_trt ../config.yaml        ../images
               <your config file>   <your test data folder>

2.1 methods:tensorrtx projcet:

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

NOTE: if you use this project ,you should have yolov4 and yolov5 project sources models to convert modlde file,so you can use the project prepared.

Go back to your project root directory

Yolov5 as a sample:

1)source activate your yolov5 conda env:

   source actviate <yolov5 conda env name>
   cd tensorrtx/yolov5

2)generate engine file --> compile yolov5.cpp

   python gen_wts.py  <input weights file >  <outputfile_name>                            
   mkdir build && cd build
   cmake ..
   make

3)run & generate yolov5x.engine and libmyplugin.so:

   [NOTE]：you should line 13 in yolov5.cpp , #define NET x  // s m l x,configure your yolov5 model  
    ./yolov5 -x 
              -s
              -l
              -m

1) test your yolov5 engine:

   sudo ./yolov5 -d  ../samples

2) mv your file

   cp -r  yolov5*.engine ../../engine_models/
   cp  -r libmyplugins.so ../../engine_models/

3.Configure your deepstream & run yolov4 and yolov5 deepstream app

In any case, the above is just to get the engine file of the model you trained. If you have another way or modify the code to generate the engine file.

Go back to your project root directory,Deepstream_Yolo/:

3.1 run Yolo Deepstream

1)compile nvdsparsebbox_Yolo.cpp ,(includes yolov4,yolov5)

   cd nvdsinfer_custom_impl_Yolo
   cmake..
   make  
   cd ..

2)configure your deepstream_app_config_yoloV.txt & onfig_infer_primary_yoloV.txt

example:

if your run yolov4,configure your deepstream_app_config_yoloV4.txt,

   deepstream-app -c deepstream_app_config_yoloV4.txt

if your use tensorrtx to generate YoloV5 enigne, add export libmyplugins.so file path,

   LD_PRELOAD=./libmyplugins.so deepstream-app -c deepstream_app_config_yoloV5.txt

Name		Name	Last commit message	Last commit date
parent directory ..
includes		includes
nvdsinfer_custom_impl_Yolo		nvdsinfer_custom_impl_Yolo
src		src
trans_project		trans_project
Lincense.txt		Lincense.txt
README.md		README.md
config_infer_primary_yoloV2.txt		config_infer_primary_yoloV2.txt
config_infer_primary_yoloV2_tiny.txt		config_infer_primary_yoloV2_tiny.txt
config_infer_primary_yoloV3.txt		config_infer_primary_yoloV3.txt
config_infer_primary_yoloV3_tiny.txt		config_infer_primary_yoloV3_tiny.txt
config_infer_primary_yoloV4.txt		config_infer_primary_yoloV4.txt
config_infer_primary_yoloV5.txt		config_infer_primary_yoloV5.txt
config_infer_primary_yolo_tlt.txt		config_infer_primary_yolo_tlt.txt
deepstream-app		deepstream-app
deepstream_app_config_yoloV2.txt		deepstream_app_config_yoloV2.txt
deepstream_app_config_yoloV2_tiny.txt		deepstream_app_config_yoloV2_tiny.txt
deepstream_app_config_yoloV3.txt		deepstream_app_config_yoloV3.txt
deepstream_app_config_yoloV3_tiny.txt		deepstream_app_config_yoloV3_tiny.txt
deepstream_app_config_yoloV4.txt		deepstream_app_config_yoloV4.txt
deepstream_app_config_yoloV5.txt		deepstream_app_config_yoloV5.txt
deepstream_app_config_yolo_tlt.txt		deepstream_app_config_yolo_tlt.txt
labels.txt		labels.txt
labels_tlt.txt		labels_tlt.txt

Files

Deepstream_Yolo

Directory actions

More options

Directory actions

More options

Latest commit

History

Deepstream_Yolo

Folders and files

parent directory

README.md

Copyright (c) 2021, positive666. All rights reserved.

0.Instruction(coming soon update )

1.Make sure your environment is configured

2.Convert the trained model file into an engine file

Here are the two ways,you can select yolov5s,yolov5m,yolov5l,yolov5x :

1.trans_project(onnx---->trt)

2.tensorrtx project

2.0 methods_1----------trans_project(onnx---->trt):

YOLOv4 (note: torch version==1.4.0)

1)darnet models--->onnx ,(your v5 and v4 need two environments,different torch version)

2)pytorch models--->onnx:

3)onnx->trt,you will get model yolov4 trt file

if you can simpilfy onnx ,make sure your onnxsim and run :

if you want to compile yolov4.cpp ,generate yolov4 trt file in your build folder:

YOLOV5(note: torch version==1.7.0 )

1)your source actvivate your python environment,input your yolov5 models into models/

2)run demo_pytorch2onnx.py to generate yolov5 onnx file:

or you can try run:

if you can simpilfy onnx ,make sure your onnxsim and run :

3)compile yolov5.cpp ,generate yolov5 enginel:

2.1 methods:tensorrtx projcet:

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

NOTE: if you use this project ,you should have yolov4 and yolov5 project sources models to convert modlde file,so you can use the project prepared.

Go back to your project root directory

Yolov5 as a sample:

1)source activate your yolov5 conda env:

2)generate engine file --> compile yolov5.cpp

3)run & generate yolov5x.engine and libmyplugin.so:

1) test your yolov5 engine:

2) mv your file

3.Configure your deepstream & run yolov4 and yolov5 deepstream app

In any case, the above is just to get the engine file of the model you trained. If you have another way or modify the code to generate the engine file.

Go back to your project root directory,Deepstream_Yolo/:

3.1 run Yolo Deepstream

1)compile nvdsparsebbox_Yolo.cpp ,(includes yolov4,yolov5)

2)configure your deepstream_app_config_yoloV.txt & onfig_infer_primary_yoloV.txt

example:

if your run yolov4,configure your deepstream_app_config_yoloV4.txt,

if your use tensorrtx to generate YoloV5 enigne, add export libmyplugins.so file path,