<h1 align="center"><span>TensorRT-YOLOv9-ROS</span></h1>
+ ROS version of [YOLOv9](https://github.com/WongKinYiu/yolov9) accelerated with [TensorRT](https://github.com/NVIDIA/TensorRT) API
+ This repository is a merely re-implementation with `ROS` of the:
+ 👏 [TensorRT-YOLOv9-C++](https://github.com/spacewalk01/TensorRT-YOLOv9), which is based on
+ [YOLOv9](https://github.com/WongKinYiu/yolov9) - `YOLOv9`: Learning What You Want to Learn Using Programmable Gradient Information.
+ [TensorRT](https://github.com/NVIDIA/TensorRT/tree/release/8.6/samples) - `TensorRT` samples and api documentation.
+ [TensorRTx](https://github.com/wang-xinyu/tensorrtx) - Implementation of popular deep learning networks with TensorRT network definition API.
<br>
### Known issues / notes
+ The resolution of image to be trained should be multiplication of 64
<br>
## Dependencies
+ `ROS` (currently supporting only `ROS1`)
+ `C++` >= 17
+ `cmake` >= 3.14
+ `OpenCV` >= 4.2
+ `TensorRT`, `CUDA`, `cuDNN`
+ `.engine` file generated with `TensorRT`
+ Tested versions:
+ Desktop with i9-10900k, RTX 3080 - `CUDA` 11.5, `cuDNN` 8.3.2.44, `TensorRT` 8.4.0.6
</details>
<br>
## You may want to:
<details><summary> ■ Unfold here to see how to install CUDA, cuDNN and TensorRT </summary>
### ● **Note that apt install with deb is preferred to run file and source file build for both of `CUDA` and `cuDNN`**
+ Download and install `CUDA` following instructions at here - https://developer.nvidia.com/cuda-downloads
+ Download and install `cuDNN` following instructions at here - https://developer.nvidia.com/cudnn-downloads
+ If you want, also refer to here - https://docs.nvidia.com/deeplearning/cudnn/installation/linux.html#
+ Set up environmental paths
```bash
gedit ~/.bashrc
*** Type and save below, CUDA_PATH should be like /usr/local/cuda-11.5, depending on your version ***
export PATH=CUDA_PATH/bin:$PATH
export LD_LIBRARY_PATH=CUDA_PATH/lib64:$LD_LIBRARY_PATH
. ~/.bashrc
gedit ~/.profile
*** Type and save below, CUDA_PATH should be like /usr/local/cuda-11.5, depending on your version ***
export PATH=CUDA_PATH/bin:$PATH
export LD_LIBRARY_PATH=CUDA_PATH/lib64:$LD_LIBRARY_PATH
. ~/.profile
```
+ Verify, if installed properly
```bash
# Verify
dpkg -l | grep cuda
dpkg -l | grep cudnn
nvcc --version
```
<br>
### ● **Note that apt install with deb is preferred to other methods for `TensorRT`**
+ Download `TensorRT` at here - https://developer.nvidia.com/tensorrt-download
+ Follow the instructions at here - https://docs.nvidia.com/deeplearning/tensorrt/install-guide/index.html#installing-debian
+ Installing full packages is recommended, which means:
```bash
sudo apt install tensorrt
sudo apt install python3-libnvinfer-dev
sudo apt install onnx-graphsurgeon
```
<br>
</details>
<details><summary> ■ Unfold here to see how to train custom data / generate TensorRT engine file with safe Python3 virtual environment </summary>
<br>
### ● Common step for training / engine file
0. Make sure that you have installed all dependencies properly.
+ Particularly, you should install full packages of `TensorRT`: `tensorrt`, `python3-libnvinfer-dev`, `onnx-graphsurgeon`
1. Install and make `Python3` virtual env
```bash
python3 -m pip install virtualenv virtualenvwrapper
cd <PATH YOU WANT TO SAVE VIRTUAL ENVIRONMENT>
virtualenv -p python3 <NAME YOU WANT>
*** Now you can activate with
source <PATH YOU SAVED>/<NAME YOU WANT>/bin/activate
*** Deactivate with
deactivate
```
2. (While virtual env being activated), clone `YOLOv9` repo and install requirements
```bash
git clone https://github.com/WongKinYiu/yolov9
cd yolov9
pip install -r requirements.txt
```
<br>
### ● Converting .pt to .onnx, and then .engine
0. (While virtual env being activated)
1. Get trained `YOLOv9` weight file as `.pt` by training your own data or downloading the pre-trained model at here - https://github.com/WongKinYiu/yolov9/releases
2. Reparameterize the `.pt` file (saving computation, memory, and size by trimming unnecessary parts for inference but necessary only for training)
```bash
cd yolov9 # cloned at above step
wget https://raw.githubusercontent.com/engcang/TensorRT_YOLOv9_ROS/main/reparameterize.py
*** Change the number of classes in the reparameterize.py in line 8 (nc=80)
python reparameterize.py yolov9-c.pt yolov9-c-reparameterized.pt # input.pt output.put
```
3. Export `.pt` file as `.onnx`
```bash
python export.py --weights yolov9-c-reparameterized.pt --include onnx
```
4. Then `.onnx` to `.engine`
```bash
/usr/src/tensorrt/bin/trtexec --onnx=yolov9-c-reparameterized.onnx --saveEngine=yolov9-c.engine
#for faster, less accurate
/usr/src/tensorrt/bin/trtexec --onnx=yolov9-c-reparameterized.onnx --saveEngine=yolov9-c-fp16.engine --fp16
#not recommended - much faster, much less accurate
/usr/src/tensorrt/bin/trtexec --onnx=yolov9-c-reparameterized.onnx --saveEngine=yolov9-c-int8.engine --int8
```
<br>
### ● Training your own data
0. (While virtual env being activated) + `YOLOv9` is cloned already, requirements are installed already
1. Prepare data and labels in `YOLO format`.
+ You may want to use this - https://github.com/AlexeyAB/Yolo_mark
+ Or `roboflow` - https://docs.ultralytics.com/yolov5/tutorials/roboflow_datasets_integration/
2. Make proper `data.yaml` file by copying and editing `yolov9/data/coco.yaml` as follows:
```yaml
path: training # dataset root dir (relative from train.py file)
train: train # train images folder (relative to 'path')
val: val # val images folder (relative to 'path')
test: test # test images folder (relative to 'path')
# Classes
names:
0: Transmission tower
1: Insulator
```
3. Make proper `yolov9.yaml` file by copying and editing `yolov9/models/detect/yolov9.yaml or yolov9-c, yolov9-e, etc.`
```yaml
# parameters
nc: 2 # number of classes
depth_multiple: 1.0 # model depth multiple
width_multiple: 1.0 # layer channel multiple
#activation: nn.LeakyReLU(0.1)
#activation: nn.ReLU()
# anchors
anchors: 3
# YOLOv9 backbone
backbone:
[
[-1, 1, Silence, []],
# conv down
[-1, 1, Conv, [64, 3, 2]], # 1-P1/2
...
]
```
4. Edit learning parameters by editing `yolov9/data/hyps/hyp.scratch-high.yaml`
5. **Put all of files properly in the `yolov9` folder. If outside the `yolov9` folder, error occurs!**
```
yolov9
│ ...
├─ data # Reference folder
│ ├─ coco.yaml
│ └─ hyps
│ └─ hyp.scratch-high.yaml
├─ models # Reference folder
│ ...
│ ├─ detect
│ ...
│ │ ├─ yolov9-c.yaml
│ │ ├─ yolov9-e.yaml
│ │ └─ yolov9.yaml
├─ runs # Output saved folder
│ ...
├─ train.py # Using this file for GELAN
├─ train_dual.py # Using this file for YOLOv9
├─ training # Using this folder
│ ├─ yolov9-c.pt
│ ├─ data.yaml
│ ├─ yolov9.yaml
│ ├─ test
│ │ ├─ 02001.jpg
│ │ ├─ 02001.txt
│ │ └─ ...
│ ├─ train
│ │ ├─ 00001.jpg
│ │ ├─ 00001.txt
│ │ └─ ...
│ ├─ val
│ │ ├─ 04000.jpg
│ │ ├─ 04000.txt
│ │ └─ ...
└─ └─ ...
```
6. Train
```bash
cd yolov9
*** Using pretrained model (yolov9-c.pt here), fine-tuning:
python train_dual.py --batch-size 4 --epochs 100 --img 640 --device 0 --close-mosaic 15 \
--data training/data.yaml --weights training/yolov9-c.pt --cfg training/yolov9.yaml --hyp data/hyps/hyp.scratch-high.yaml
*** From the scratch:
python train_dual.py --batch-size 4 --epochs 100 --img 640 --device 0 --close-mosaic 15 \
--data training/data.yaml --weights '' --cfg training/yolov9.yaml --hyp data/hyps/hyp.scratch-high.yaml
```
<br>
### ● Trouble shooting for training
0. (While virtual env being activated)
1. `AttributeError: 'FreeTypeFont' object has no attribute 'getsize'`
+ This is because installed Pillow version is too recent.
+ Solve with `pip install Pillow==9.5.0`
2. Getting `Killed
没有合适的资源?快使用搜索试试~ 我知道了~
算法部署-基于ROS+TensorRT+Cpp部署YOLOv9目标检测算法.zip
共16个文件
h:5个
msg:2个
yaml:1个
1.该资源内容由用户上传,如若侵权请联系客服进行举报
2.虚拟产品一经售出概不退款(资源遇到问题,请及时私信上传者)
2.虚拟产品一经售出概不退款(资源遇到问题,请及时私信上传者)
版权申诉
0 下载量 67 浏览量
2024-10-18
16:09:23
上传
评论 1
收藏 23KB ZIP 举报
温馨提示
算法部署_基于ROS+TensorRT+Cpp部署YOLOv9目标检测算法
资源推荐
资源详情
资源评论
收起资源包目录
算法部署_基于ROS+TensorRT+Cpp部署YOLOv9目标检测算法.zip (16个子文件)
算法部署_基于ROS+TensorRT+Cpp部署YOLOv9目标检测算法
include
macros.h 420B
preprocess.h 308B
cuda_utils.h 1KB
preprocess.cu 4KB
logging.h 16KB
yolov9.hpp 8KB
CMakeLists.txt 1KB
src
main.cpp 479B
main.h 6KB
package.xml 938B
reparameterize.py 4KB
launch
run.launch 243B
msg
bboxes.msg 37B
bbox.msg 77B
README.md 9KB
config
config.yaml 1KB
共 16 条
- 1
资源评论
__AtYou__
- 粉丝: 3512
- 资源: 2175
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
最新资源
资源上传下载、课程学习等过程中有任何疑问或建议,欢迎提出宝贵意见哦~我们会及时处理!
点击此处反馈
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功